Skip to content

D
DeepSpeech

PaddlePaddle / DeepSpeech
大约 2 年 前同步成功

通知 210
Star 8425
Fork 1598
D
DeepSpeech
提交 598eb1a5 编写于 作者: T tianhao zhang
浏览文件
操作

Merge branch 'develop' into fix_dp_init

上级 9560d650 0975a332
展开全部 隐藏空白更改
内联 并排
Showing with 3892 addition and 1314 deletion
CHANGELOG.md 已删除 100644 → 0
浏览文件 @ 9560d650
# Changelog
Date: 2022-3-22, Author: yt605155624.
Add features to: CLI:
- Support aishell3_hifigan、vctk_hifigan
- PRLink: https://github.com/PaddlePaddle/PaddleSpeech/pull/1587
Date: 2022-3-09, Author: yt605155624.
Add features to: T2S:
- Add ljspeech hifigan egs.
- PRLink: https://github.com/PaddlePaddle/PaddleSpeech/pull/1549
Date: 2022-3-08, Author: yt605155624.
Add features to: T2S:
- Add aishell3 hifigan egs.
- PRLink: https://github.com/PaddlePaddle/PaddleSpeech/pull/1545
Date: 2022-3-08, Author: yt605155624.
Add features to: T2S:
- Add vctk hifigan egs.
- PRLink: https://github.com/PaddlePaddle/PaddleSpeech/pull/1544
Date: 2022-1-29, Author: yt605155624.
Add features to: T2S:
- Update aishell3 vc0 with new Tacotron2.
- PRLink: https://github.com/PaddlePaddle/PaddleSpeech/pull/1419
Date: 2022-1-29, Author: yt605155624.
Add features to: T2S:
- Add ljspeech Tacotron2.
- PRLink: https://github.com/PaddlePaddle/PaddleSpeech/pull/1416
Date: 2022-1-24, Author: yt605155624.
Add features to: T2S:
- Add csmsc WaveRNN.
- PRLink: https://github.com/PaddlePaddle/PaddleSpeech/pull/1379
Date: 2022-1-19, Author: yt605155624.
Add features to: T2S:
- Add csmsc Tacotron2.
- PRLink: https://github.com/PaddlePaddle/PaddleSpeech/pull/1314
Date: 2022-1-10, Author: Jackwaterveg.
Add features to: CLI:
- Support English (librispeech/asr1/transformer).
- Support choosing `decode_method` for conformer and transformer models.
- Refactor the config, using the unified config.
- PRLink: https://github.com/PaddlePaddle/PaddleSpeech/pull/1297
***
Date: 2022-1-17, Author: Jackwaterveg.
Add features to: CLI:
- Support deepspeech2 online/offline model(aishell).
- PRLink: https://github.com/PaddlePaddle/PaddleSpeech/pull/1356
***
Date: 2022-1-24, Author: Jackwaterveg.
Add features to: ctc_decoders:
- Support online ctc prefix-beam search decoder.
- Unified ctc online decoder and ctc offline decoder.
- PRLink: https://github.com/PaddlePaddle/PaddleSpeech/pull/821
***
README.md
浏览文件 @ 598eb1a5
......@@ -159,15 +159,20 @@ Via the easy-to-use, efficient, flexible and scalable implementation, our vision
- 🧩 *Cascaded models application*: as an extension of the typical traditional audio tasks, we combine the workflows of the aforementioned tasks with other fields like Natural language processing (NLP) and Computer Vision (CV).
### Recent Update
- 👑 2022.05.13: Release [PP-ASR](./docs/source/asr/PPASR.md)、[PP-TTS](./docs/source/tts/PPTTS.md)、[PP-VPR](docs/source/vpr/PPVPR.md)
- 👏🏻 2022.05.06: `Streaming ASR` with `Punctuation Restoration` and `Token Timestamp`.
- 👏🏻 2022.05.06: `Server` is available for `Speaker Verification`, and `Punctuation Restoration`.
- 👏🏻 2022.04.28: `Streaming Server` is available for `Automatic Speech Recognition` and `Text-to-Speech`.
- 👏🏻 2022.03.28: `Server` is available for `Audio Classification`, `Automatic Speech Recognition` and `Text-to-Speech`.
- 👏🏻 2022.03.28: `CLI` is available for `Speaker Verification`.
- ⚡ 2022.08.25: Release TTS [finetune](./examples/other/tts_finetune/tts3) example.
- 🔥 2022.08.22: Add ERNIE-SAT models: [ERNIE-SAT-vctk](./examples/vctk/ernie_sat)、[ERNIE-SAT-aishell3](./examples/aishell3/ernie_sat)、[ERNIE-SAT-zh_en](./examples/aishell3_vctk/ernie_sat).
- 🔥 2022.08.15: Add [g2pW](https://github.com/GitYCC/g2pW) into TTS Chinese Text Frontend.
- 🔥 2022.08.09: Release [Chinese English mixed TTS](./examples/zh_en_tts/tts3).
- ⚡ 2022.08.03: Add ONNXRuntime infer for TTS CLI.
- 🎉 2022.07.18: Release VITS: [VITS-csmsc](./examples/csmsc/vits)、[VITS-aishell3](./examples/aishell3/vits)、[VITS-VC](./examples/aishell3/vits-vc).
- 🎉 2022.06.22: All TTS models support ONNX format.
- 🍀 2022.06.17: Add [PaddleSpeech Web Demo](./demos/speech_web).
- 👑 2022.05.13: Release [PP-ASR](./docs/source/asr/PPASR.md)、[PP-TTS](./docs/source/tts/PPTTS.md)、[PP-VPR](docs/source/vpr/PPVPR.md).
- 👏🏻 2022.05.06: `PaddleSpeech Streaming Server` is available for `Streaming ASR` with `Punctuation Restoration` and `Token Timestamp` and `Text-to-Speech`.
- 👏🏻 2022.05.06: `PaddleSpeech Server` is available for `Audio Classification`, `Automatic Speech Recognition` and `Text-to-Speech`, `Speaker Verification` and `Punctuation Restoration`.
- 👏🏻 2022.03.28: `PaddleSpeech CLI` is available for `Speaker Verification`.
- 🤗 2021.12.14: [ASR](https://huggingface.co/spaces/KPatrick/PaddleSpeechASR) and [TTS](https://huggingface.co/spaces/KPatrick/PaddleSpeechTTS) Demos on Hugging Face Spaces are available!
- 👏🏻 2021.12.10: `CLI` is available for `Audio Classification`, `Automatic Speech Recognition`, `Speech Translation (English to Chinese)` and `Text-to-Speech`.
- 👏🏻 2021.12.10: `PaddleSpeech CLI` is available for `Audio Classification`, `Automatic Speech Recognition`, `Speech Translation (English to Chinese)` and `Text-to-Speech`.
### Community
- Scan the QR code below with your Wechat, you can access to official technical exchange group and get the bonus ( more than 20GB learning materials, such as papers, codes and videos ) and the live link of the lessons. Look forward to your participation.
......@@ -376,7 +381,7 @@ Developers can have a try of our speech server with [PaddleSpeech Server Command
**Start server**
```shell
paddlespeech_server start --config_file ./paddlespeech/server/conf/application.yaml
paddlespeech_server start --config_file ./demos/speech_server/conf/application.yaml
```
**Access Speech Recognition Services**
......@@ -599,47 +604,61 @@ PaddleSpeech supports a series of most popular models. They are summarized in [r
</td>
</tr>
<tr>
<td >HiFiGAN</td>
<td >LJSpeech / VCTK / CSMSC / AISHELL-3</td>
<td>HiFiGAN</td>
<td>LJSpeech / VCTK / CSMSC / AISHELL-3</td>
<td>
<a href = "./examples/ljspeech/voc5">HiFiGAN-ljspeech</a> / <a href = "./examples/vctk/voc5">HiFiGAN-vctk</a> / <a href = "./examples/csmsc/voc5">HiFiGAN-csmsc</a> / <a href = "./examples/aishell3/voc5">HiFiGAN-aishell3</a>
</td>
</tr>
<tr>
<td >WaveRNN</td>
<td >CSMSC</td>
<td>WaveRNN</td>
<td>CSMSC</td>
<td>
<a href = "./examples/csmsc/voc6">WaveRNN-csmsc</a>
</td>
</tr>
<tr>
<td rowspan="3">Voice Cloning</td>
<td rowspan="5">Voice Cloning</td>
<td>GE2E</td>
<td >Librispeech, etc.</td>
<td>
<a href = "./examples/other/ge2e">ge2e</a>
<a href = "./examples/other/ge2e">GE2E</a>
</td>
</tr>
<tr>
<td>SV2TTS (GE2E + Tacotron2)</td>
<td>AISHELL-3</td>
<td>
<a href = "./examples/aishell3/vc0">VC0</a>
</td>
</tr>
<tr>
<td>SV2TTS (GE2E + FastSpeech2)</td>
<td>AISHELL-3</td>
<td>
<a href = "./examples/aishell3/vc1">VC1</a>
</td>
</tr>
<tr>
<td>GE2E + Tacotron2</td>
<td>SV2TTS (ECAPA-TDNN + FastSpeech2)</td>
<td>AISHELL-3</td>
<td>
<a href = "./examples/aishell3/vc0">ge2e-tacotron2-aishell3</a>
<a href = "./examples/aishell3/vc2">VC2</a>
</td>
</tr>
<tr>
<td>GE2E + FastSpeech2</td>
<td>GE2E + VITS</td>
<td>AISHELL-3</td>
<td>
<a href = "./examples/aishell3/vc1">ge2e-fastspeech2-aishell3</a>
<a href = "./examples/aishell3/vits-vc">VITS-VC</a>
</td>
</tr>
<tr>
<tr>
<td rowspan="3">End-to-End</td>
<td>VITS</td>
<td >CSMSC</td>
<td>CSMSC / AISHELL-3</td>
<td>
<a href = "./examples/csmsc/vits">VITS-csmsc</a>
<a href = "./examples/csmsc/vits">VITS-csmsc</a> / <a href = "./examples/aishell3/vits">VITS-aishell3</a>
</td>
</tr>
</tbody>
......@@ -869,8 +888,9 @@ You are warmly welcome to submit questions in [discussions](https://github.com/P
</p>
## Acknowledgement
- Many thanks to [david-95](https://github.com/david-95) improved TTS, fixed multi-punctuation bug, and contributed to multiple program and data.
- Many thanks to [BarryKCL](https://github.com/BarryKCL) improved TTS Chinses frontend based on [G2PW](https://github.com/GitYCC/g2pW)
- Many thanks to [HighCWu](https://github.com/HighCWu) for adding [VITS-aishell3](./examples/aishell3/vits) and [VITS-VC](./examples/aishell3/vits-vc) examples.
- Many thanks to [david-95](https://github.com/david-95) improved TTS, fixed multi-punctuation bug, and contributed to multiple program and data.
- Many thanks to [BarryKCL](https://github.com/BarryKCL) improved TTS Chinses frontend based on [G2PW](https://github.com/GitYCC/g2pW).
- Many thanks to [yeyupiaoling](https://github.com/yeyupiaoling)/[PPASR](https://github.com/yeyupiaoling/PPASR)/[PaddlePaddle-DeepSpeech](https://github.com/yeyupiaoling/PaddlePaddle-DeepSpeech)/[VoiceprintRecognition-PaddlePaddle](https://github.com/yeyupiaoling/VoiceprintRecognition-PaddlePaddle)/[AudioClassification-PaddlePaddle](https://github.com/yeyupiaoling/AudioClassification-PaddlePaddle) for years of attention, constructive advice and great help.
- Many thanks to [mymagicpower](https://github.com/mymagicpower) for the Java implementation of ASR upon [short](https://github.com/mymagicpower/AIAS/tree/main/3_audio_sdks/asr_sdk) and [long](https://github.com/mymagicpower/AIAS/tree/main/3_audio_sdks/asr_long_audio_sdk) audio files.
- Many thanks to [JiehangXie](https://github.com/JiehangXie)/[PaddleBoBo](https://github.com/JiehangXie/PaddleBoBo) for developing Virtual Uploader(VUP)/Virtual YouTuber(VTuber) with PaddleSpeech TTS function.
......
README_cn.md
浏览文件 @ 598eb1a5
......@@ -164,13 +164,37 @@
- 🧩 级联模型应用: 作为传统语音任务的扩展，我们结合了自然语言处理、计算机视觉等任务，实现更接近实际需求的产业级应用。
### 近期更新
### 近期活动
❗️重磅❗️飞桨智慧金融行业系列直播课
✅ 覆盖智能风控、智能运维、智能营销、智能客服四大金融主流场景
📆 9月6日-9月29日每周二、四19:00
+ 智慧金融行业深入洞察
+ 8节理论+实践精品直播课
+ 10+真实产业场景范例教学及实践
+ 更有免费算力+结业证书等礼品等你来拿
扫码报名码住直播链接，与行业精英深度交流
<div align="center">
<img src="https://user-images.githubusercontent.com/30135920/188431897-a02f028f-dd13-41e8-8ff6-749468cdc850.jpg" width = "200" />
</div>
### 近期更新
- ⚡ 2022.08.25: 发布 TTS [finetune](./examples/other/tts_finetune/tts3) 示例。
- 🔥 2022.08.22: 新增 ERNIE-SAT 模型: [ERNIE-SAT-vctk](./examples/vctk/ernie_sat)、[ERNIE-SAT-aishell3](./examples/aishell3/ernie_sat)、[ERNIE-SAT-zh_en](./examples/aishell3_vctk/ernie_sat)。
- 🔥 2022.08.15: 将 [g2pW](https://github.com/GitYCC/g2pW) 引入 TTS 中文文本前端。
- 🔥 2022.08.09: 发布[中英文混合 TTS](./examples/zh_en_tts/tts3)。
- ⚡ 2022.08.03: TTS CLI 新增 ONNXRuntime 推理方式。
- 🎉 2022.07.18: 发布 VITS 模型: [VITS-csmsc](./examples/csmsc/vits)、[VITS-aishell3](./examples/aishell3/vits)、[VITS-VC](./examples/aishell3/vits-vc)。
- 🎉 2022.06.22: 所有 TTS 模型支持了 ONNX 格式。
- 🍀 2022.06.17: 新增 [PaddleSpeech 网页应用](./demos/speech_web)。
- 👑 2022.05.13: PaddleSpeech 发布 [PP-ASR](./docs/source/asr/PPASR_cn.md) 流式语音识别系统、[PP-TTS](./docs/source/tts/PPTTS_cn.md) 流式语音合成系统、[PP-VPR](docs/source/vpr/PPVPR_cn.md) 全链路声纹识别系统
- 👏🏻 2022.05.06: PaddleSpeech Streaming Server 上线! 覆盖了语音识别（标点恢复、时间戳），和语音合成。
- 👏🏻 2022.05.06: PaddleSpeech Server 上线! 覆盖了声音分类、语音识别、语音合成、声纹识别，标点恢复。
- 👏🏻 2022.03.28: PaddleSpeech CLI 覆盖声音分类、语音识别、语音翻译（英译中）、语音合成，声纹验证。
- 🤗 2021.12.14: PaddleSpeech [ASR](https://huggingface.co/spaces/KPatrick/PaddleSpeechASR) and [TTS](https://huggingface.co/spaces/KPatrick/PaddleSpeechTTS) Demos on Hugging Face Spaces are available!
- 👏🏻 2022.05.06: PaddleSpeech Streaming Server 上线！覆盖了语音识别（标点恢复、时间戳）和语音合成。
- 👏🏻 2022.05.06: PaddleSpeech Server 上线！覆盖了声音分类、语音识别、语音合成、声纹识别，标点恢复。
- 👏🏻 2022.03.28: PaddleSpeech CLI 覆盖声音分类、语音识别、语音翻译（英译中）、语音合成和声纹验证。
- 🤗 2021.12.14: PaddleSpeech [ASR](https://huggingface.co/spaces/KPatrick/PaddleSpeechASR) 和 [TTS](https://huggingface.co/spaces/KPatrick/PaddleSpeechTTS) 可在 Hugging Face Spaces 上体验！
- 👏🏻 2021.12.10: PaddleSpeech CLI 支持语音分类, 语音识别, 语音翻译（英译中）和语音合成。
### 🔥 加入技术交流群获取入群福利
......@@ -221,7 +245,6 @@ pip install .
<a name="快速开始"></a>
## 快速开始
安装完成后，开发者可以通过命令行或者 Python 快速开始，命令行模式下改变 `--input` 可以尝试用自己的音频或文本测试，支持 16k wav 格式音频。
你也可以在 `aistudio` 中快速体验 👉🏻[一键预测，快速上手 Speech 开发任务](https://aistudio.baidu.com/aistudio/projectdetail/4353348?sUid=2470186&shared=1&ts=1660878142250)。
......@@ -377,7 +400,7 @@ python API 一键预测
**启动服务**
```shell
paddlespeech_server start --config_file ./paddlespeech/server/conf/application.yaml
paddlespeech_server start --config_file ./demos/speech_server/conf/application.yaml
```
**访问语音识别服务**
......@@ -608,34 +631,47 @@ PaddleSpeech 的 **语音合成** 主要包含三个模块：文本前端、声
</td>
</tr>
<tr>
<td rowspan="3">声音克隆</td>
<td rowspan="5">声音克隆</td>
<td>GE2E</td>
<td >Librispeech, etc.</td>
<td>
<a href = "./examples/other/ge2e">ge2e</a>
<a href = "./examples/other/ge2e">GE2E</a>
</td>
</tr>
<tr>
<td>SV2TTS (GE2E + Tacotron2)</td>
<td>AISHELL-3</td>
<td>
<a href = "./examples/aishell3/vc0">VC0</a>
</td>
</tr>
<tr>
<td>GE2E + Tacotron2</td>
<td>SV2TTS (GE2E + FastSpeech2)</td>
<td>AISHELL-3</td>
<td>
<a href = "./examples/aishell3/vc0">ge2e-tacotron2-aishell3</a>
<a href = "./examples/aishell3/vc1">VC1</a>
</td>
</tr>
<tr>
<td>GE2E + FastSpeech2</td>
<td>SV2TTS (ECAPA-TDNN + FastSpeech2)</td>
<td>AISHELL-3</td>
<td>
<a href = "./examples/aishell3/vc1">ge2e-fastspeech2-aishell3</a>
<a href = "./examples/aishell3/vc2">VC2</a>
</td>
</tr>
<tr>
<td>GE2E + VITS</td>
<td>AISHELL-3</td>
<td>
<a href = "./examples/aishell3/vits-vc">VITS-VC</a>
</td>
</tr>
<tr>
<td rowspan="3">端到端</td>
<td>VITS</td>
<td >CSMSC</td>
<td>CSMSC / AISHELL-3</td>
<td>
<a href = "./examples/csmsc/vits">VITS-csmsc</a>
<a href = "./examples/csmsc/vits">VITS-csmsc</a> / <a href = "./examples/aishell3/vits">VITS-aishell3</a>
</td>
</tr>
</tbody>
......@@ -873,8 +909,9 @@ PaddleSpeech 的 **语音合成** 主要包含三个模块：文本前端、声
</p>
## 致谢
- 非常感谢 [david-95](https://github.com/david-95)修复句尾多标点符号出错的问题，补充frontend语音polyphonic 数据，贡献补充多条程序和数据
- 非常感谢 [BarryKCL](https://github.com/BarryKCL)基于[G2PW](https://github.com/GitYCC/g2pW)对TTS中文文本前端的优化。
- 非常感谢 [HighCWu](https://github.com/HighCWu) 新增 [VITS-aishell3](./examples/aishell3/vits) 和 [VITS-VC](./examples/aishell3/vits-vc) 代码示例。
- 非常感谢 [david-95](https://github.com/david-95) 修复句尾多标点符号出错的问题，贡献补充多条程序和数据。
- 非常感谢 [BarryKCL](https://github.com/BarryKCL) 基于 [G2PW](https://github.com/GitYCC/g2pW) 对 TTS 中文文本前端的优化。
- 非常感谢 [yeyupiaoling](https://github.com/yeyupiaoling)/[PPASR](https://github.com/yeyupiaoling/PPASR)/[PaddlePaddle-DeepSpeech](https://github.com/yeyupiaoling/PaddlePaddle-DeepSpeech)/[VoiceprintRecognition-PaddlePaddle](https://github.com/yeyupiaoling/VoiceprintRecognition-PaddlePaddle)/[AudioClassification-PaddlePaddle](https://github.com/yeyupiaoling/AudioClassification-PaddlePaddle) 多年来的关注和建议，以及在诸多问题上的帮助。
- 非常感谢 [mymagicpower](https://github.com/mymagicpower) 采用PaddleSpeech 对 ASR 的[短语音](https://github.com/mymagicpower/AIAS/tree/main/3_audio_sdks/asr_sdk)及[长语音](https://github.com/mymagicpower/AIAS/tree/main/3_audio_sdks/asr_long_audio_sdk)进行 Java 实现。
- 非常感谢 [JiehangXie](https://github.com/JiehangXie)/[PaddleBoBo](https://github.com/JiehangXie/PaddleBoBo) 采用 PaddleSpeech 语音合成功能实现 Virtual Uploader(VUP)/Virtual YouTuber(VTuber) 虚拟主播。
......
demos/audio_searching/README.md
浏览文件 @ 598eb1a5
......@@ -226,6 +226,12 @@ recall and elapsed time statistics are shown in the following figure：
The retrieval framework based on Milvus takes about 2.9 milliseconds to retrieve on the premise of 90% recall rate, and it takes about 500 milliseconds for feature extraction (testing audio takes about 5 seconds), that is, a single audio test takes about 503 milliseconds in total, which can meet most application scenarios.
* compute embeding takes 500 ms
* retrieval with cosine takes 2.9 ms
* total takes 503 ms
> test audio is 5 sec
### 6.Pretrained Models
Here is a list of pretrained models released by PaddleSpeech :
......
demos/audio_searching/src/operations/load.py
浏览文件 @ 598eb1a5
......@@ -26,8 +26,9 @@ def get_audios(path):
"""
supported_formats = [".wav", ".mp3", ".ogg", ".flac", ".m4a"]
return [
item for sublist in [[os.path.join(dir, file) for file in files]
for dir, _, files in list(os.walk(path))]
item
for sublist in [[os.path.join(dir, file) for file in files]
for dir, _, files in list(os.walk(path))]
for item in sublist if os.path.splitext(item)[1] in supported_formats
]
......
demos/metaverse/README.md
浏览文件 @ 598eb1a5
([简体中文](./README_cn.md)|English)
# Metaverse
## Introduction
Metaverse is a new Internet application and social form integrating virtual reality produced by integrating a variety of new technologies.
......
demos/metaverse/README_cn.md 0 → 100644
浏览文件 @ 598eb1a5
(简体中文|[English](./README.md))
# Metaverse
## 简介
Metaverse 是一种新的互联网应用和社交形式，融合了多种新技术，产生了虚拟现实。
这个演示是一个让图片中的名人“说话”的实现。通过 `PaddleSpeech` 的 `TTS` 模块和 `PaddleGAN` 的组合，我们集成了安装和特定模块到一个 shell 脚本中。
## 使用
您可以使用 `PaddleSpeech` 的 `TTS` 模块和 `PaddleGAN` 让您最喜欢的人说出指定的内容，并构建您的虚拟人。
运行 `run.sh` 完成所有基本程序，包括安装。
```bash
./run.sh
```
在 `run.sh`, 先会执行 `source path.sh` 来设置好环境变量。
如果您想尝试您的句子，请替换 `sentences.txt` 中的句子。
如果您想尝试图像，请将图像替换 shell 脚本中的 `download/Lamarr.png` 。
结果已显示在我们的 [notebook](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/docs/tutorial/tts/tts_tutorial.ipynb)。
demos/speaker_verification/README.md
浏览文件 @ 598eb1a5
......@@ -19,6 +19,7 @@ The input of this cli demo should be a WAV file(`.wav`), and the sample rate mus
Here are sample files for this demo that can be downloaded:
```bash
wget -c https://paddlespeech.bj.bcebos.com/vector/audio/85236145389.wav
wget -c https://paddlespeech.bj.bcebos.com/vector/audio/123456789.wav
```
### 3. Usage
......
demos/speaker_verification/README_cn.md
浏览文件 @ 598eb1a5
......@@ -19,6 +19,7 @@
```bash
# 该音频的内容是数字串 85236145389
wget -c https://paddlespeech.bj.bcebos.com/vector/audio/85236145389.wav
wget -c https://paddlespeech.bj.bcebos.com/vector/audio/123456789.wav
```
### 3. 使用方法
- 命令行 (推荐使用)
......
demos/speech_web/API.md
浏览文件 @ 598eb1a5
......@@ -401,4 +401,4 @@ curl -X 'GET' \
"code": 0,
"result":"AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA",
"message": "ok"
```
\ No newline at end of file
```
demos/speech_web/speech_server/main.py
浏览文件 @ 598eb1a5
......@@ -3,48 +3,48 @@
# 2. 接收录音音频，返回识别结果
# 3. 接收ASR识别结果，返回NLP对话结果
# 4. 接收NLP对话结果，返回TTS音频
import argparse
import base64
import yaml
import os
import json
import datetime
import json
import os
from typing import List
import aiofiles
import librosa
import soundfile as sf
import numpy as np
import argparse
import uvicorn
import aiofiles
from typing import Optional, List
from pydantic import BaseModel
from fastapi import FastAPI, Header, File, UploadFile, Form, Cookie, WebSocket, WebSocketDisconnect
from fastapi import FastAPI
from fastapi import File
from fastapi import Form
from fastapi import UploadFile
from fastapi import WebSocket
from fastapi import WebSocketDisconnect
from fastapi.responses import StreamingResponse
from starlette.responses import FileResponse
from starlette.middleware.cors import CORSMiddleware
from starlette.requests import Request
from starlette.websockets import WebSocketState as WebSocketState
from pydantic import BaseModel
from src.AudioManeger import AudioMannger
from src.util import *
from src.robot import Robot
from src.WebsocketManeger import ConnectionManager
from src.SpeechBase.vpr import VPR
from src.util import *
from src.WebsocketManeger import ConnectionManager
from starlette.middleware.cors import CORSMiddleware
from starlette.requests import Request
from starlette.responses import FileResponse
from starlette.websockets import WebSocketState as WebSocketState
from paddlespeech.server.engine.asr.online.python.asr_engine import PaddleASRConnectionHanddler
from paddlespeech.server.utils.audio_process import float2pcm
# 解析配置
parser = argparse.ArgumentParser(
prog='PaddleSpeechDemo', add_help=True)
parser = argparse.ArgumentParser(prog='PaddleSpeechDemo', add_help=True)
parser.add_argument(
"--port",
action="store",
type=int,
help="port of the app",
default=8010,
required=False)
"--port",
action="store",
type=int,
help="port of the app",
default=8010,
required=False)
args = parser.parse_args()
port = args.port
......@@ -60,39 +60,41 @@ ie_model_path = "source/model"
UPLOAD_PATH = "source/vpr"
WAV_PATH = "source/wav"
base_sources = [
UPLOAD_PATH, WAV_PATH
]
base_sources = [UPLOAD_PATH, WAV_PATH]
for path in base_sources:
os.makedirs(path, exist_ok=True)
# 初始化
app = FastAPI()
chatbot = Robot(asr_config, tts_config, asr_init_path, ie_model_path=ie_model_path)
chatbot = Robot(
asr_config, tts_config, asr_init_path, ie_model_path=ie_model_path)
manager = ConnectionManager()
aumanager = AudioMannger(chatbot)
aumanager.init()
vpr = VPR(db_path, dim = 192, top_k = 5)
vpr = VPR(db_path, dim=192, top_k=5)
# 服务配置
class NlpBase(BaseModel):
chat: str
class TtsBase(BaseModel):
text: str
text: str
class Audios:
def __init__(self) -> None:
self.audios = b""
audios = Audios()
######################################################################
########################### ASR 服务 #################################
#####################################################################
# 接收文件，返回ASR结果
# 上传文件
@app.post("/asr/offline")
......@@ -101,7 +103,8 @@ async def speech2textOffline(files: List[UploadFile]):
asr_res = ""
for file in files[:1]:
# 生成时间戳
now_name = "asr_offline_" + datetime.datetime.strftime(datetime.datetime.now(), '%Y%m%d%H%M%S') + randName() + ".wav"
now_name = "asr_offline_" + datetime.datetime.strftime(
datetime.datetime.now(), '%Y%m%d%H%M%S') + randName() + ".wav"
out_file_path = os.path.join(WAV_PATH, now_name)
async with aiofiles.open(out_file_path, 'wb') as out_file:
content = await file.read() # async read
......@@ -110,10 +113,9 @@ async def speech2textOffline(files: List[UploadFile]):
# 返回ASR识别结果
asr_res = chatbot.speech2text(out_file_path)
return SuccessRequest(result=asr_res)
# else:
# return ErrorRequest(message="文件不是.wav格式")
return ErrorRequest(message="上传文件为空")
# 接收文件，同时将wav强制转成16k, int16类型
@app.post("/asr/offlinefile")
async def speech2textOfflineFile(files: List[UploadFile]):
......@@ -121,7 +123,8 @@ async def speech2textOfflineFile(files: List[UploadFile]):
asr_res = ""
for file in files[:1]:
# 生成时间戳
now_name = "asr_offline_" + datetime.datetime.strftime(datetime.datetime.now(), '%Y%m%d%H%M%S') + randName() + ".wav"
now_name = "asr_offline_" + datetime.datetime.strftime(
datetime.datetime.now(), '%Y%m%d%H%M%S') + randName() + ".wav"
out_file_path = os.path.join(WAV_PATH, now_name)
async with aiofiles.open(out_file_path, 'wb') as out_file:
content = await file.read() # async read
......@@ -132,22 +135,18 @@ async def speech2textOfflineFile(files: List[UploadFile]):
wav = float2pcm(wav) # float32 to int16
wav_bytes = wav.tobytes() # to bytes
wav_base64 = base64.b64encode(wav_bytes).decode('utf8')
# 将文件重新写入
now_name = now_name[:-4] + "_16k" + ".wav"
out_file_path = os.path.join(WAV_PATH, now_name)
sf.write(out_file_path,wav,16000)
sf.write(out_file_path, wav, 16000)
# 返回ASR识别结果
asr_res = chatbot.speech2text(out_file_path)
response_res = {
"asr_result": asr_res,
"wav_base64": wav_base64
}
response_res = {"asr_result": asr_res, "wav_base64": wav_base64}
return SuccessRequest(result=response_res)
return ErrorRequest(message="上传文件为空")
return ErrorRequest(message="上传文件为空")
# 流式接收测试
......@@ -161,15 +160,17 @@ async def speech2textOnlineRecive(files: List[UploadFile]):
print(f"audios长度变化: {len(audios.audios)}")
return SuccessRequest(message="接收成功")
# 采集环境噪音大小
@app.post("/asr/collectEnv")
async def collectEnv(files: List[UploadFile]):
for file in files[:1]:
for file in files[:1]:
content = await file.read() # async read
# 初始化, wav 前44字节是头部信息
aumanager.compute_env_volume(content[44:])
vad_ = aumanager.vad_threshold
return SuccessRequest(result=vad_,message="采集环境噪音成功")
return SuccessRequest(result=vad_, message="采集环境噪音成功")
# 停止录音
@app.get("/asr/stopRecord")
......@@ -179,6 +180,7 @@ async def stopRecord():
print("Online录音暂停")
return SuccessRequest(message="停止成功")
# 恢复录音
@app.get("/asr/resumeRecord")
async def resumeRecord():
......@@ -187,7 +189,7 @@ async def resumeRecord():
return SuccessRequest(message="Online录音恢复")
# 聊天用的ASR
# 聊天用的 ASR
@app.websocket("/ws/asr/offlineStream")
async def websocket_endpoint(websocket: WebSocket):
await manager.connect(websocket)
......@@ -210,9 +212,9 @@ async def websocket_endpoint(websocket: WebSocket):
# print(f"用户-{user}-离开")
# Online识别的ASR
# 流式识别的 ASR
@app.websocket('/ws/asr/onlineStream')
async def websocket_endpoint(websocket: WebSocket):
async def websocket_endpoint_online(websocket: WebSocket):
"""PaddleSpeech Online ASR Server api
Args:
......@@ -298,12 +300,14 @@ async def websocket_endpoint(websocket: WebSocket):
except WebSocketDisconnect:
pass
######################################################################
########################### NLP 服务 #################################
#####################################################################
@app.post("/nlp/chat")
async def chatOffline(nlp_base:NlpBase):
async def chatOffline(nlp_base: NlpBase):
chat = nlp_base.chat
if not chat:
return ErrorRequest(message="传入文本为空")
......@@ -311,8 +315,9 @@ async def chatOffline(nlp_base:NlpBase):
res = chatbot.chat(chat)
return SuccessRequest(result=res)
@app.post("/nlp/ie")
async def ieOffline(nlp_base:NlpBase):
async def ieOffline(nlp_base: NlpBase):
nlp_text = nlp_base.chat
if not nlp_text:
return ErrorRequest(message="传入文本为空")
......@@ -320,17 +325,20 @@ async def ieOffline(nlp_base:NlpBase):
res = chatbot.ie(nlp_text)
return SuccessRequest(result=res)
######################################################################
########################### TTS 服务 #################################
#####################################################################
@app.post("/tts/offline")
async def text2speechOffline(tts_base:TtsBase):
async def text2speechOffline(tts_base: TtsBase):
text = tts_base.text
if not text:
return ErrorRequest(message="文本为空")
else:
now_name = "tts_"+ datetime.datetime.strftime(datetime.datetime.now(), '%Y%m%d%H%M%S') + randName() + ".wav"
now_name = "tts_" + datetime.datetime.strftime(
datetime.datetime.now(), '%Y%m%d%H%M%S') + randName() + ".wav"
out_file_path = os.path.join(WAV_PATH, now_name)
# 保存为文件，再转成base64传输
chatbot.text2speech(text, outpath=out_file_path)
......@@ -339,12 +347,14 @@ async def text2speechOffline(tts_base:TtsBase):
base_str = base64.b64encode(data_bin)
return SuccessRequest(result=base_str)
# http流式TTS
@app.post("/tts/online")
async def stream_tts(request_body: TtsBase):
text = request_body.text
return StreamingResponse(chatbot.text2speechStreamBytes(text=text))
# ws流式TTS
@app.websocket("/ws/tts/online")
async def stream_ttsWS(websocket: WebSocket):
......@@ -356,17 +366,11 @@ async def stream_ttsWS(websocket: WebSocket):
if text:
for sub_wav in chatbot.text2speechStream(text=text):
# print("发送sub wav: ", len(sub_wav))
res = {
"wav": sub_wav,
"done": False
}
res = {"wav": sub_wav, "done": False}
await websocket.send_json(res)
# 输送结束
res = {
"wav": sub_wav,
"done": True
}
res = {"wav": sub_wav, "done": True}
await websocket.send_json(res)
# manager.disconnect(websocket)
......@@ -396,8 +400,9 @@ async def vpr_enroll(table_name: str=None,
return {'status': False, 'msg': "spk_id can not be None"}
# Save the upload data to server.
content = await audio.read()
now_name = "vpr_enroll_" + datetime.datetime.strftime(datetime.datetime.now(), '%Y%m%d%H%M%S') + randName() + ".wav"
audio_path = os.path.join(UPLOAD_PATH, now_name)
now_name = "vpr_enroll_" + datetime.datetime.strftime(
datetime.datetime.now(), '%Y%m%d%H%M%S') + randName() + ".wav"
audio_path = os.path.join(UPLOAD_PATH, now_name)
with open(audio_path, "wb+") as f:
f.write(content)
......@@ -413,20 +418,19 @@ async def vpr_recog(request: Request,
audio: UploadFile=File(...)):
# Voice print recognition online
# try:
# Save the upload data to server.
# Save the upload data to server.
content = await audio.read()
now_name = "vpr_query_" + datetime.datetime.strftime(datetime.datetime.now(), '%Y%m%d%H%M%S') + randName() + ".wav"
query_audio_path = os.path.join(UPLOAD_PATH, now_name)
now_name = "vpr_query_" + datetime.datetime.strftime(
datetime.datetime.now(), '%Y%m%d%H%M%S') + randName() + ".wav"
query_audio_path = os.path.join(UPLOAD_PATH, now_name)
with open(query_audio_path, "wb+") as f:
f.write(content)
f.write(content)
spk_ids, paths, scores = vpr.do_search_vpr(query_audio_path)
res = dict(zip(spk_ids, zip(paths, scores)))
# Sort results by distance metric, closest distances first
res = sorted(res.items(), key=lambda item: item[1][1], reverse=True)
return res
# except Exception as e:
# return {'status': False, 'msg': e}, 400
@app.post('/vpr/del')
......@@ -460,17 +464,18 @@ async def vpr_database64(vprId: int):
return {'status': False, 'msg': "vpr_id can not be None"}
audio_path = vpr.do_get_wav(vprId)
# 返回base64
# 将文件转成16k, 16bit类型的wav文件
wav, sr = librosa.load(audio_path, sr=16000)
wav = float2pcm(wav) # float32 to int16
wav_bytes = wav.tobytes() # to bytes
wav_base64 = base64.b64encode(wav_bytes).decode('utf8')
return SuccessRequest(result=wav_base64)
except Exception as e:
return {'status': False, 'msg': e}, 400
@app.get('/vpr/data')
async def vpr_data(vprId: int):
# Get the audio file from path by spk_id in MySQL
......@@ -482,11 +487,6 @@ async def vpr_data(vprId: int):
except Exception as e:
return {'status': False, 'msg': e}, 400
if __name__ == '__main__':
uvicorn.run(app=app, host='0.0.0.0', port=port)
demos/speech_web/speech_server/requirements.txt
浏览文件 @ 598eb1a5
aiofiles
faiss-cpu
fastapi
librosa
numpy
paddlenlp
paddlepaddle
paddlespeech
pydantic
scikit_learn
python-multipartscikit_learn
SoundFile
starlette
uvicorn
paddlepaddle
paddlespeech
paddlenlp
faiss-cpu
python-multipart
\ No newline at end of file
demos/speech_web/speech_server/src/AudioManeger.py
浏览文件 @ 598eb1a5
import imp
from queue import Queue
import numpy as np
import datetime
import os
import wave
import random
import datetime
import numpy as np
from .util import randName
class AudioMannger:
def __init__(self, robot, frame_length=160, frame=10, data_width=2, vad_default = 300):
def __init__(self,
robot,
frame_length=160,
frame=10,
data_width=2,
vad_default=300):
# 二进制 pcm 流
self.audios = b''
self.asr_result = ""
......@@ -20,8 +24,9 @@ class AudioMannger:
os.makedirs(self.file_dir, exist_ok=True)
self.vad_deafult = vad_default
self.vad_threshold = vad_default
self.vad_threshold_path = os.path.join(self.file_dir, "vad_threshold.npy")
self.vad_threshold_path = os.path.join(self.file_dir,
"vad_threshold.npy")
# 10ms 一帧
self.frame_length = frame_length
# 10帧，检测一次 vad
......@@ -30,67 +35,64 @@ class AudioMannger:
self.data_width = data_width
# window
self.window_length = frame_length * frame * data_width
# 是否开始录音
self.on_asr = False
self.silence_cnt = 0
self.silence_cnt = 0
self.max_silence_cnt = 4
self.is_pause = False # 录音暂停与恢复
def init(self):
if os.path.exists(self.vad_threshold_path):
# 平均响度文件存在
self.vad_threshold = np.load(self.vad_threshold_path)
def clear_audio(self):
# 清空 pcm 累积片段与 asr 识别结果
self.audios = b''
def clear_asr(self):
self.asr_result = ""
def compute_chunk_volume(self, start_index, pcm_bins):
# 根据帧长计算能量平均值
pcm_bin = pcm_bins[start_index: start_index + self.window_length]
pcm_bin = pcm_bins[start_index:start_index + self.window_length]
# 转成 numpy
pcm_np = np.frombuffer(pcm_bin, np.int16)
# 归一化 + 计算响度
x = pcm_np.astype(np.float32)
x = np.abs(x)
return np.mean(x)
return np.mean(x)
def is_speech(self, start_index, pcm_bins):
# 检查是否没
if start_index > len(pcm_bins):
return False
# 检查从这个 start 开始是否为静音帧
energy = self.compute_chunk_volume(start_index=start_index, pcm_bins=pcm_bins)
energy = self.compute_chunk_volume(
start_index=start_index, pcm_bins=pcm_bins)
# print(energy)
if energy > self.vad_threshold:
return True
else:
return False
def compute_env_volume(self, pcm_bins):
max_energy = 0
start = 0
while start < len(pcm_bins):
energy = self.compute_chunk_volume(start_index=start, pcm_bins=pcm_bins)
energy = self.compute_chunk_volume(
start_index=start, pcm_bins=pcm_bins)
if energy > max_energy:
max_energy = energy
start += self.window_length
self.vad_threshold = max_energy + 100 if max_energy > self.vad_deafult else self.vad_deafult
# 保存成文件
np.save(self.vad_threshold_path, self.vad_threshold)
print(f"vad 阈值大小: {self.vad_threshold}")
print(f"环境采样保存: {os.path.realpath(self.vad_threshold_path)}")
def stream_asr(self, pcm_bin):
# 先把 pcm_bin 送进去做端点检测
start = 0
......@@ -99,7 +101,7 @@ class AudioMannger:
self.on_asr = True
self.silence_cnt = 0
print("录音中")
self.audios += pcm_bin[ start : start + self.window_length]
self.audios += pcm_bin[start:start + self.window_length]
else:
if self.on_asr:
self.silence_cnt += 1
......@@ -110,41 +112,42 @@ class AudioMannger:
print("录音停止")
# audios 保存为 wav, 送入 ASR
if len(self.audios) > 2 * 16000:
file_path = os.path.join(self.file_dir, "asr_" + datetime.datetime.strftime(datetime.datetime.now(), '%Y%m%d%H%M%S') + randName() + ".wav")
file_path = os.path.join(
self.file_dir,
"asr_" + datetime.datetime.strftime(
datetime.datetime.now(),
'%Y%m%d%H%M%S') + randName() + ".wav")
self.save_audio(file_path=file_path)
self.asr_result = self.robot.speech2text(file_path)
self.clear_audio()
return self.asr_result
return self.asr_result
else:
# 正常接收
print("录音中 静音")
self.audios += pcm_bin[ start : start + self.window_length]
self.audios += pcm_bin[start:start + self.window_length]
start += self.window_length
return ""
def save_audio(self, file_path):
print("保存音频")
wf = wave.open(file_path, 'wb') # 创建一个音频文件，名字为“01.wav"
wf.setnchannels(1) # 设置声道数为2
wf.setsampwidth(2) # 设置采样深度为
wf.setframerate(16000) # 设置采样率为16000
wf = wave.open(file_path, 'wb') # 创建一个音频文件，名字为“01.wav"
wf.setnchannels(1) # 设置声道数为2
wf.setsampwidth(2) # 设置采样深度为
wf.setframerate(16000) # 设置采样率为16000
# 将数据写入创建的音频文件
wf.writeframes(self.audios)
# 写完后将文件关闭
wf.close()
def end(self):
# audios 保存为 wav, 送入 ASR
file_path = os.path.join(self.file_dir, "asr.wav")
self.save_audio(file_path=file_path)
return self.robot.speech2text(file_path)
def stop(self):
self.is_pause = True
self.audios = b''
def resume(self):
self.is_pause = False
\ No newline at end of file
demos/speech_web/speech_server/src/SpeechBase/asr.py
浏览文件 @ 598eb1a5
from re import sub
import numpy as np
import paddle
import librosa
import soundfile
from paddlespeech.server.engine.asr.online.python.asr_engine import ASREngine
from paddlespeech.server.engine.asr.online.python.asr_engine import PaddleASRConnectionHanddler
from paddlespeech.server.utils.config import get_config
def readWave(samples):
x_len = len(samples)
......@@ -31,20 +28,23 @@ def readWave(samples):
class ASR:
def __init__(self, config_path, ) -> None:
def __init__(
self,
config_path, ) -> None:
self.config = get_config(config_path)['asr_online']
self.engine = ASREngine()
self.engine.init(self.config)
self.connection_handler = PaddleASRConnectionHanddler(self.engine)
def offlineASR(self, samples, sample_rate=16000):
x_chunk, x_chunk_lens = self.engine.preprocess(samples=samples, sample_rate=sample_rate)
x_chunk, x_chunk_lens = self.engine.preprocess(
samples=samples, sample_rate=sample_rate)
self.engine.run(x_chunk, x_chunk_lens)
result = self.engine.postprocess()
self.engine.reset()
return result
def onlineASR(self, samples:bytes=None, is_finished=False):
def onlineASR(self, samples: bytes=None, is_finished=False):
if not is_finished:
# 流式开始
self.connection_handler.extract_feat(samples)
......@@ -58,5 +58,3 @@ class ASR:
asr_results = self.connection_handler.get_result()
self.connection_handler.reset()
return asr_results
\ No newline at end of file
demos/speech_web/speech_server/src/SpeechBase/nlp.py
浏览文件 @ 598eb1a5
from paddlenlp import Taskflow
class NLP:
def __init__(self, ie_model_path=None):
schema = ["时间", "出发地", "目的地", "费用"]
if ie_model_path:
self.ie_model = Taskflow("information_extraction",
schema=schema, task_path=ie_model_path)
self.ie_model = Taskflow(
"information_extraction",
schema=schema,
task_path=ie_model_path)
else:
self.ie_model = Taskflow("information_extraction",
schema=schema)
self.ie_model = Taskflow("information_extraction", schema=schema)
self.dialogue_model = Taskflow("dialogue")
def chat(self, text):
result = self.dialogue_model([text])
return result[0]
def ie(self, text):
result = self.ie_model(text)
return result
\ No newline at end of file
demos/speech_web/speech_server/src/SpeechBase/sql_helper.py
浏览文件 @ 598eb1a5
import base64
import sqlite3
import os
import sqlite3
import numpy as np
from pkg_resources import resource_stream
def dict_factory(cursor, row):
d = {}
for idx, col in enumerate(cursor.description):
d[col[0]] = row[idx]
return d
def dict_factory(cursor, row):
d = {}
for idx, col in enumerate(cursor.description):
d[col[0]] = row[idx]
return d
class DataBase(object):
def __init__(self, db_path:str):
def __init__(self, db_path: str):
db_path = os.path.realpath(db_path)
if os.path.exists(db_path):
......@@ -21,12 +22,12 @@ class DataBase(object):
db_path_dir = os.path.dirname(db_path)
os.makedirs(db_path_dir, exist_ok=True)
self.db_path = db_path
self.conn = sqlite3.connect(self.db_path)
self.conn.row_factory = dict_factory
self.cursor = self.conn.cursor()
self.init_database()
def init_database(self):
"""
初始化数据库， 若表不存在则创建
......@@ -41,20 +42,21 @@ class DataBase(object):
"""
self.cursor.execute(sql)
self.conn.commit()
def execute_base(self, sql, data_dict):
self.cursor.execute(sql, data_dict)
self.conn.commit()
def insert_one(self, username, vector_base64:str, wav_path):
def insert_one(self, username, vector_base64: str, wav_path):
if not os.path.exists(wav_path):
return None, "wav not exists"
else:
sql = f"""
sql = """
insert into
vprtable (username, vector, wavpath)
values (?, ?, ?)
"""
try:
self.cursor.execute(sql, (username, vector_base64, wav_path))
self.conn.commit()
......@@ -63,25 +65,27 @@ class DataBase(object):
except Exception as e:
print(e)
return None, e
def select_all(self):
sql = """
SELECT * from vprtable
"""
result = self.cursor.execute(sql).fetchall()
return result
def select_by_id(self, vpr_id):
sql = f"""
SELECT * from vprtable WHERE `id` = {vpr_id}
"""
result = self.cursor.execute(sql).fetchall()
return result
def select_by_username(self, username):
sql = f"""
SELECT * from vprtable WHERE `username` = '{username}'
"""
result = self.cursor.execute(sql).fetchall()
return result
......@@ -89,28 +93,30 @@ class DataBase(object):
sql = f"""
DELETE from vprtable WHERE `username`='{username}'
"""
self.cursor.execute(sql)
self.conn.commit()
def drop_all(self):
sql = f"""
sql = """
DELETE from vprtable
"""
self.cursor.execute(sql)
self.conn.commit()
def drop_table(self):
sql = f"""
sql = """
DROP TABLE vprtable
"""
self.cursor.execute(sql)
self.conn.commit()
def encode_vector(self, vector:np.ndarray):
def encode_vector(self, vector: np.ndarray):
return base64.b64encode(vector).decode('utf8')
def decode_vector(self, vector_base64, dtype=np.float32):
b = base64.b64decode(vector_base64)
vc = np.frombuffer(b, dtype=dtype)
return vc
\ No newline at end of file
demos/speech_web/speech_server/src/SpeechBase/tts.py
浏览文件 @ 598eb1a5
......@@ -5,18 +5,19 @@
# 2. 加载模型
# 3. 端到端推理
# 4. 流式推理
import base64
import math
import logging
import math
import numpy as np
from paddlespeech.server.utils.onnx_infer import get_sess
from paddlespeech.t2s.frontend.zh_frontend import Frontend
from paddlespeech.server.utils.util import denorm, get_chunks
from paddlespeech.server.engine.tts.online.onnx.tts_engine import TTSEngine
from paddlespeech.server.utils.audio_process import float2pcm
from paddlespeech.server.utils.config import get_config
from paddlespeech.server.utils.util import denorm
from paddlespeech.server.utils.util import get_chunks
from paddlespeech.t2s.frontend.zh_frontend import Frontend
from paddlespeech.server.engine.tts.online.onnx.tts_engine import TTSEngine
class TTS:
def __init__(self, config_path):
......@@ -26,12 +27,12 @@ class TTS:
self.engine.init(self.config)
self.executor = self.engine.executor
#self.engine.warm_up()
# 前端初始化
self.frontend = Frontend(
phone_vocab_path=self.engine.executor.phones_dict,
tone_vocab_path=None)
phone_vocab_path=self.engine.executor.phones_dict,
tone_vocab_path=None)
def depadding(self, data, chunk_num, chunk_id, block, pad, upsample):
"""
Streaming inference removes the result of pad inference
......@@ -48,39 +49,37 @@ class TTS:
data = data[front_pad * upsample:(front_pad + block) * upsample]
return data
def offlineTTS(self, text):
get_tone_ids = False
merge_sentences = False
input_ids = self.frontend.get_input_ids(
text,
merge_sentences=merge_sentences,
get_tone_ids=get_tone_ids)
text, merge_sentences=merge_sentences, get_tone_ids=get_tone_ids)
phone_ids = input_ids["phone_ids"]
wav_list = []
for i in range(len(phone_ids)):
orig_hs = self.engine.executor.am_encoder_infer_sess.run(
None, input_feed={'text': phone_ids[i].numpy()}
)
None, input_feed={'text': phone_ids[i].numpy()})
hs = orig_hs[0]
am_decoder_output = self.engine.executor.am_decoder_sess.run(
None, input_feed={'xs': hs})
None, input_feed={'xs': hs})
am_postnet_output = self.engine.executor.am_postnet_sess.run(
None,
input_feed={
'xs': np.transpose(am_decoder_output[0], (0, 2, 1))
})
None,
input_feed={
'xs': np.transpose(am_decoder_output[0], (0, 2, 1))
})
am_output_data = am_decoder_output + np.transpose(
am_postnet_output[0], (0, 2, 1))
normalized_mel = am_output_data[0][0]
mel = denorm(normalized_mel, self.engine.executor.am_mu, self.engine.executor.am_std)
mel = denorm(normalized_mel, self.engine.executor.am_mu,
self.engine.executor.am_std)
wav = self.engine.executor.voc_sess.run(
output_names=None, input_feed={'logmel': mel})[0]
output_names=None, input_feed={'logmel': mel})[0]
wav_list.append(wav)
wavs = np.concatenate(wav_list)
return wavs
def streamTTS(self, text):
get_tone_ids = False
......@@ -88,9 +87,7 @@ class TTS:
# front
input_ids = self.frontend.get_input_ids(
text,
merge_sentences=merge_sentences,
get_tone_ids=get_tone_ids)
text, merge_sentences=merge_sentences, get_tone_ids=get_tone_ids)
phone_ids = input_ids["phone_ids"]
for i in range(len(phone_ids)):
......@@ -105,14 +102,15 @@ class TTS:
mel = mel[0]
# voc streaming
mel_chunks = get_chunks(mel, self.config.voc_block, self.config.voc_pad, "voc")
mel_chunks = get_chunks(mel, self.config.voc_block,
self.config.voc_pad, "voc")
voc_chunk_num = len(mel_chunks)
for i, mel_chunk in enumerate(mel_chunks):
sub_wav = self.executor.voc_sess.run(
output_names=None, input_feed={'logmel': mel_chunk})
sub_wav = self.depadding(sub_wav[0], voc_chunk_num, i,
self.config.voc_block, self.config.voc_pad,
self.config.voc_upsample)
sub_wav = self.depadding(
sub_wav[0], voc_chunk_num, i, self.config.voc_block,
self.config.voc_pad, self.config.voc_upsample)
yield self.after_process(sub_wav)
......@@ -130,7 +128,8 @@ class TTS:
end = min(self.config.voc_block + self.config.voc_pad, mel_len)
# streaming am
hss = get_chunks(orig_hs, self.config.am_block, self.config.am_pad, "am")
hss = get_chunks(orig_hs, self.config.am_block,
self.config.am_pad, "am")
am_chunk_num = len(hss)
for i, hs in enumerate(hss):
am_decoder_output = self.executor.am_decoder_sess.run(
......@@ -147,7 +146,8 @@ class TTS:
sub_mel = denorm(normalized_mel, self.executor.am_mu,
self.executor.am_std)
sub_mel = self.depadding(sub_mel, am_chunk_num, i,
self.config.am_block, self.config.am_pad, 1)
self.config.am_block,
self.config.am_pad, 1)
if i == 0:
mel_streaming = sub_mel
......@@ -165,23 +165,22 @@ class TTS:
output_names=None, input_feed={'logmel': voc_chunk})
sub_wav = self.depadding(
sub_wav[0], voc_chunk_num, voc_chunk_id,
self.config.voc_block, self.config.voc_pad, self.config.voc_upsample)
self.config.voc_block, self.config.voc_pad,
self.config.voc_upsample)
yield self.after_process(sub_wav)
voc_chunk_id += 1
start = max(
0, voc_chunk_id * self.config.voc_block - self.config.voc_pad)
end = min(
(voc_chunk_id + 1) * self.config.voc_block + self.config.voc_pad,
mel_len)
start = max(0, voc_chunk_id * self.config.voc_block -
self.config.voc_pad)
end = min((voc_chunk_id + 1) * self.config.voc_block +
self.config.voc_pad, mel_len)
else:
logging.error(
"Only support fastspeech2_csmsc or fastspeech2_cnndecoder_csmsc on streaming tts."
)
)
def streamTTSBytes(self, text):
for wav in self.engine.executor.infer(
text=text,
......@@ -191,19 +190,14 @@ class TTS:
wav = float2pcm(wav) # float32 to int16
wav_bytes = wav.tobytes() # to bytes
yield wav_bytes
def after_process(self, wav):
# for tvm
wav = float2pcm(wav) # float32 to int16
wav_bytes = wav.tobytes() # to bytes
wav_base64 = base64.b64encode(wav_bytes).decode('utf8') # to base64
return wav_base64
def streamTTS_TVM(self, text):
# 用 TVM 优化
pass
\ No newline at end of file
demos/speech_web/speech_server/src/SpeechBase/vpr.py
浏览文件 @ 598eb1a5
# vpr Demo 没有使用 mysql 与 muilvs, 仅用于docker演示
import logging
import faiss
from matplotlib import use
import numpy as np
from .sql_helper import DataBase
from .vpr_encode import get_audio_embedding
class VPR:
def __init__(self, db_path, dim, top_k) -> None:
# 初始化
......@@ -14,15 +16,15 @@ class VPR:
self.top_k = top_k
self.dtype = np.float32
self.vpr_idx = 0
# db 初始化
self.db = DataBase(db_path)
# faiss 初始化
index_ip = faiss.IndexFlatIP(dim)
self.index_ip = faiss.IndexIDMap(index_ip)
self.init()
def init(self):
# demo 初始化，把 mysql中的向量注册到 faiss 中
sql_dbs = self.db.select_all()
......@@ -34,12 +36,13 @@ class VPR:
if len(vc.shape) == 1:
vc = np.expand_dims(vc, axis=0)
# 构建数据库
self.index_ip.add_with_ids(vc, np.array((idx,)).astype('int64'))
self.index_ip.add_with_ids(vc, np.array(
(idx, )).astype('int64'))
logging.info("faiss 构建完毕")
def faiss_enroll(self, idx, vc):
self.index_ip.add_with_ids(vc, np.array((idx,)).astype('int64'))
self.index_ip.add_with_ids(vc, np.array((idx, )).astype('int64'))
def vpr_enroll(self, username, wav_path):
# 注册声纹
emb = get_audio_embedding(wav_path)
......@@ -53,21 +56,22 @@ class VPR:
else:
last_idx, mess = None
return last_idx
def vpr_recog(self, wav_path):
# 识别声纹
emb_search = get_audio_embedding(wav_path)
if emb_search is not None:
emb_search = np.expand_dims(emb_search, axis=0)
D, I = self.index_ip.search(emb_search, self.top_k)
D = D.tolist()[0]
I = I.tolist()[0]
return [(round(D[i] * 100, 2 ), I[i]) for i in range(len(D)) if I[i] != -1]
I = I.tolist()[0]
return [(round(D[i] * 100, 2), I[i]) for i in range(len(D))
if I[i] != -1]
else:
logging.error("识别失败")
return None
def do_search_vpr(self, wav_path):
spk_ids, paths, scores = [], [], []
recog_result = self.vpr_recog(wav_path)
......@@ -78,41 +82,39 @@ class VPR:
scores.append(score)
paths.append("")
return spk_ids, paths, scores
def vpr_del(self, username):
# 根据用户username, 删除声纹
# 查用户ID，删除对应向量
res = self.db.select_by_username(username)
for r in res:
idx = r['id']
self.index_ip.remove_ids(np.array((idx,)).astype('int64'))
self.index_ip.remove_ids(np.array((idx, )).astype('int64'))
self.db.drop_by_username(username)
def vpr_list(self):
# 获取数据列表
return self.db.select_all()
def do_list(self):
spk_ids, vpr_ids = [], []
for res in self.db.select_all():
spk_ids.append(res['username'])
vpr_ids.append(res['id'])
return spk_ids, vpr_ids
return spk_ids, vpr_ids
def do_get_wav(self, vpr_idx):
res = self.db.select_by_id(vpr_idx)
return res[0]['wavpath']
res = self.db.select_by_id(vpr_idx)
return res[0]['wavpath']
def vpr_data(self, idx):
# 获取对应ID的数据
res = self.db.select_by_id(idx)
return res
def vpr_droptable(self):
# 删除表
self.db.drop_table()
# 清空 faiss
self.index_ip.reset()
demos/speech_web/speech_server/src/SpeechBase/vpr_encode.py
浏览文件 @ 598eb1a5
from paddlespeech.cli.vector import VectorExecutor
import numpy as np
import logging
import numpy as np
from paddlespeech.cli.vector import VectorExecutor
vector_executor = VectorExecutor()
def get_audio_embedding(path):
"""
Use vpr_inference to generate embedding of audio
......@@ -16,5 +19,3 @@ def get_audio_embedding(path):
except Exception as e:
logging.error(f"Error with embedding:{e}")
return None
\ No newline at end of file
demos/speech_web/speech_server/src/WebsocketManeger.py
浏览文件 @ 598eb1a5
......@@ -2,6 +2,7 @@ from typing import List
from fastapi import WebSocket
class ConnectionManager:
def __init__(self):
# 存放激活的ws连接对象
......@@ -28,4 +29,4 @@ class ConnectionManager:
await connection.send_text(message)
manager = ConnectionManager()
\ No newline at end of file
manager = ConnectionManager()
demos/speech_web/speech_server/src/robot.py
浏览文件 @ 598eb1a5
from paddlespeech.cli.asr.infer import ASRExecutor
import soundfile as sf
import os
import librosa
import soundfile as sf
from src.SpeechBase.asr import ASR
from src.SpeechBase.tts import TTS
from src.SpeechBase.nlp import NLP
from src.SpeechBase.tts import TTS
from paddlespeech.cli.asr.infer import ASRExecutor
class Robot:
def __init__(self, asr_config, tts_config,asr_init_path,
def __init__(self,
asr_config,
tts_config,
asr_init_path,
ie_model_path=None) -> None:
self.nlp = NLP(ie_model_path=ie_model_path)
self.asr = ASR(config_path=asr_config)
self.tts = TTS(config_path=tts_config)
self.tts_sample_rate = 24000
self.asr_sample_rate = 16000
# 流式识别效果不如端到端的模型，这里流式模型与端到端模型分开
self.asr_model = ASRExecutor()
self.asr_name = "conformer_wenetspeech"
self.warm_up_asrmodel(asr_init_path)
def warm_up_asrmodel(self, asr_init_path):
def warm_up_asrmodel(self, asr_init_path):
if not os.path.exists(asr_init_path):
path_dir = os.path.dirname(asr_init_path)
if not os.path.exists(path_dir):
os.makedirs(path_dir, exist_ok=True)
# TTS生成，采样率24000
text = "生成初始音频"
self.text2speech(text, asr_init_path)
# asr model初始化
self.asr_model(asr_init_path, model=self.asr_name,lang='zh',
sample_rate=16000, force_yes=True)
self.asr_model(
asr_init_path,
model=self.asr_name,
lang='zh',
sample_rate=16000,
force_yes=True)
def speech2text(self, audio_file):
self.asr_model.preprocess(self.asr_name, audio_file)
self.asr_model.infer(self.asr_name)
res = self.asr_model.postprocess()
return res
def text2speech(self, text, outpath):
wav = self.tts.offlineTTS(text)
sf.write(
outpath, wav, samplerate=self.tts_sample_rate)
sf.write(outpath, wav, samplerate=self.tts_sample_rate)
res = wav
return res
def text2speechStream(self, text):
for sub_wav_base64 in self.tts.streamTTS(text=text):
yield sub_wav_base64
def text2speechStreamBytes(self, text):
for wav_bytes in self.tts.streamTTSBytes(text=text):
yield wav_bytes
......@@ -66,5 +70,3 @@ class Robot:
def ie(self, text):
result = self.nlp.ie(text)
return result
\ No newline at end of file
demos/speech_web/speech_server/src/util.py
浏览文件 @ 598eb1a5
import random
def randName(n=5):
return "".join(random.sample('zyxwvutsrqponmlkjihgfedcba',n))
return "".join(random.sample('zyxwvutsrqponmlkjihgfedcba', n))
def SuccessRequest(result=None, message="ok"):
return {
"code": 0,
"result":result,
"message": message
}
return {"code": 0, "result": result, "message": message}
def ErrorRequest(result=None, message="error"):
return {
"code": -1,
"result":result,
"message": message
}
\ No newline at end of file
return {"code": -1, "result": result, "message": message}
demos/story_talker/README.md
浏览文件 @ 598eb1a5
([简体中文](./README_cn.md)|English)
# Story Talker
## Introduction
Storybooks are very important children's enlightenment books, but parents usually don't have enough time to read storybooks for their children. For very young children, they may not understand the Chinese characters in storybooks. Or sometimes, children just want to "listen" but don't want to "read".
......
demos/story_talker/README_cn.md 0 → 100644
浏览文件 @ 598eb1a5
(简体中文|[English](./README.md))
# Story Talker
## 简介
故事书是非常重要的儿童启蒙书，但家长通常没有足够的时间为孩子读故事书。对于非常小的孩子，他们可能不理解故事书中的汉字。或有时，孩子们只是想“听”，而不想“读”。
您可以使用 `PaddleOCR` 获取故事书的文本，并通过 `PaddleSpeech` 的 `TTS` 模块进行阅读。
## 使用
运行以下命令行开始：
```
./run.sh
```
结果已显示在 [notebook](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/docs/tutorial/tts/tts_tutorial.ipynb)。
demos/streaming_asr_server/conf/application.yaml
浏览文件 @ 598eb1a5
......@@ -28,6 +28,7 @@ asr_online:
sample_rate: 16000
cfg_path:
decode_method:
num_decoding_left_chunks: -1
force_yes: True
device: 'cpu' # cpu or gpu:id
decode_method: "attention_rescoring"
......
demos/streaming_asr_server/local/rtf_from_log.py
浏览文件 @ 598eb1a5
......@@ -34,7 +34,7 @@ if __name__ == '__main__':
n = 0
for m in rtfs:
# not accurate, may have duplicate log
n += 1
n += 1
T += m['T']
P += m['P']
......
demos/style_fs2/README.md
浏览文件 @ 598eb1a5
([简体中文](./README_cn.md)|English)
# Style FastSpeech2
## Introduction
[FastSpeech2](https://arxiv.org/abs/2006.04558) is a classical acoustic model for Text-to-Speech synthesis, which introduces controllable speech input, including `phoneme duration`、 `energy` and `pitch`.
......
demos/style_fs2/README_cn.md 0 → 100644
浏览文件 @ 598eb1a5
(简体中文|[English](./README.md))
# Style FastSpeech2
## 简介
[FastSpeech2](https://arxiv.org/abs/2006.04558) 是用于语音合成的经典声学模型，它引入了可控语音输入，包括 `phoneme duration` 、 `energy` 和 `pitch` 。
在预测阶段，您可以更改这些变量以获得一些有趣的结果。
例如:
1. `FastSpeech2` 中的 `duration` 可以控制音频的速度 ，并保持 `pitch` 。（在某些语音工具中，增加速度将增加音调，反之亦然。）
2. 当我们将一个句子的 `pitch` 设置为平均值并将音素的 `tones` 设置为 `1` 时，我们将获得 `robot-style` 的音色。
3. 当我们提高成年女性的 `pitch` （比例固定）时，我们会得到 `child-style` 的音色。
句子中不同音素的 `duration` 和 `pitch` 可以具有不同的比例。您可以设置不同的音阶比例来强调或削弱某些音素的发音。
## 运行
运行以下命令行开始：
```
./run.sh
```
在 `run.sh`, 会首先执行 `source path.sh` 去设置好环境变量。
如果您想尝试您的句子，请替换 `sentences.txt`中的句子。
更多的细节，请查看 `style_syn.py`。
语音样例可以在 [style-control-in-fastspeech2](https://paddlespeech.readthedocs.io/en/latest/tts/demo.html#style-control-in-fastspeech2) 查看。
demos/text_to_speech/README.md
浏览文件 @ 598eb1a5
......@@ -16,8 +16,8 @@ You can choose one way from easy, meduim and hard to install paddlespeech.
The input of this demo should be a text of the specific language that can be passed via argument.
### 3. Usage
- Command Line (Recommended)
The default acoustic model is `Fastspeech2`, and the default vocoder is `HiFiGAN`, the default inference method is dygraph inference.
- Chinese
The default acoustic model is `Fastspeech2`, and the default vocoder is `Parallel WaveGAN`.
```bash
paddlespeech tts --input "你好，欢迎使用百度飞桨深度学习框架！"
```
......@@ -58,6 +58,20 @@ The input of this demo should be a text of the specific language that can be pas
paddlespeech tts --am fastspeech2_mix --voc pwgan_csmsc --lang mix --input "我们的声学模型使用了 Fast Speech Two, 声码器使用了 Parallel Wave GAN and Hifi GAN." --spk_id 175 --output mix_spk175_pwgan.wav
paddlespeech tts --am fastspeech2_mix --voc hifigan_csmsc --lang mix --input "我们的声学模型使用了 Fast Speech Two, 声码器使用了 Parallel Wave GAN and Hifi GAN." --spk_id 175 --output mix_spk175.wav
```
- Use ONNXRuntime infer：
```bash
paddlespeech tts --input "你好，欢迎使用百度飞桨深度学习框架！" --output default.wav --use_onnx True
paddlespeech tts --am speedyspeech_csmsc --input "你好，欢迎使用百度飞桨深度学习框架！" --output ss.wav --use_onnx True
paddlespeech tts --voc mb_melgan_csmsc --input "你好，欢迎使用百度飞桨深度学习框架！" --output mb.wav --use_onnx True
paddlespeech tts --voc pwgan_csmsc --input "你好，欢迎使用百度飞桨深度学习框架！" --output pwgan.wav --use_onnx True
paddlespeech tts --am fastspeech2_aishell3 --voc pwgan_aishell3 --input "你好，欢迎使用百度飞桨深度学习框架！" --spk_id 0 --output aishell3_fs2_pwgan.wav --use_onnx True
paddlespeech tts --am fastspeech2_aishell3 --voc hifigan_aishell3 --input "你好，欢迎使用百度飞桨深度学习框架！" --spk_id 0 --output aishell3_fs2_hifigan.wav --use_onnx True
paddlespeech tts --am fastspeech2_ljspeech --voc pwgan_ljspeech --lang en --input "Life was like a box of chocolates, you never know what you're gonna get." --output lj_fs2_pwgan.wav --use_onnx True
paddlespeech tts --am fastspeech2_ljspeech --voc hifigan_ljspeech --lang en --input "Life was like a box of chocolates, you never know what you're gonna get." --output lj_fs2_hifigan.wav --use_onnx True
paddlespeech tts --am fastspeech2_vctk --voc pwgan_vctk --input "Life was like a box of chocolates, you never know what you're gonna get." --lang en --spk_id 0 --output vctk_fs2_pwgan.wav --use_onnx True
paddlespeech tts --am fastspeech2_vctk --voc hifigan_vctk --input "Life was like a box of chocolates, you never know what you're gonna get." --lang en --spk_id 0 --output vctk_fs2_hifigan.wav --use_onnx True
```
Usage:
```bash
......@@ -80,6 +94,8 @@ The input of this demo should be a text of the specific language that can be pas
- `lang`: Language of tts task. Default: `zh`.
- `device`: Choose device to execute model inference. Default: default device of paddlepaddle in current environment.
- `output`: Output wave filepath. Default: `output.wav`.
- `use_onnx`: whether to usen ONNXRuntime inference.
- `fs`: sample rate for ONNX models when use specified model files.
Output:
```bash
......@@ -87,38 +103,50 @@ The input of this demo should be a text of the specific language that can be pas
```
- Python API
```python
import paddle
from paddlespeech.cli.tts import TTSExecutor
tts_executor = TTSExecutor()
wav_file = tts_executor(
text='今天的天气不错啊',
output='output.wav',
am='fastspeech2_csmsc',
am_config=None,
am_ckpt=None,
am_stat=None,
spk_id=0,
phones_dict=None,
tones_dict=None,
speaker_dict=None,
voc='pwgan_csmsc',
voc_config=None,
voc_ckpt=None,
voc_stat=None,
lang='zh',
device=paddle.get_device())
print('Wave file has been generated: {}'.format(wav_file))
```
- Dygraph infer:
```python
import paddle
from paddlespeech.cli.tts import TTSExecutor
tts_executor = TTSExecutor()
wav_file = tts_executor(
text='今天的天气不错啊',
output='output.wav',
am='fastspeech2_csmsc',
am_config=None,
am_ckpt=None,
am_stat=None,
spk_id=0,
phones_dict=None,
tones_dict=None,
speaker_dict=None,
voc='pwgan_csmsc',
voc_config=None,
voc_ckpt=None,
voc_stat=None,
lang='zh',
device=paddle.get_device())
print('Wave file has been generated: {}'.format(wav_file))
```
- ONNXRuntime infer:
```python
from paddlespeech.cli.tts import TTSExecutor
tts_executor = TTSExecutor()
wav_file = tts_executor(
text='对数据集进行预处理',
output='output.wav',
am='fastspeech2_csmsc',
voc='hifigan_csmsc',
lang='zh',
use_onnx=True,
cpu_threads=2)
```
Output:
```bash
Wave file has been generated: output.wav
```
### 4. Pretrained Models
Here is a list of pretrained models released by PaddleSpeech that can be used by command and python API:
- Acoustic model
......
demos/text_to_speech/README_cn.md
浏览文件 @ 598eb1a5
(简体中文|[English](./README.md))
# 语音合成
## 介绍
语音合成是一种自然语言建模过程，其将文本转换为语音以进行音频演示。
这个 demo 是一个从给定文本生成音频的实现，它可以通过使用 `PaddleSpeech` 的单个命令或 python 中的几行代码来实现。
## 使用方法
### 1. 安装
请看[安装文档](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/docs/source/install_cn.md)。
你可以从 easy，medium，hard 三中方式中选择一种方式安装。
你可以从 easy，medium，hard 三种方式中选择一种方式安装。
### 2. 准备输入
这个 demo 的输入是通过参数传递的特定语言的文本。
### 3. 使用方法
- 命令行 (推荐使用)
默认的声学模型是 `Fastspeech2`，默认的声码器是 `HiFiGAN`，默认推理方式是动态图推理。
- 中文
默认的声学模型是 `Fastspeech2`，默认的声码器是 `Parallel WaveGAN`.
```bash
paddlespeech tts --input "你好，欢迎使用百度飞桨深度学习框架！"
```
......@@ -61,6 +58,19 @@
paddlespeech tts --am fastspeech2_mix --voc pwgan_csmsc --lang mix --input "我们的声学模型使用了 Fast Speech Two, 声码器使用了 Parallel Wave GAN and Hifi GAN." --spk_id 175 --output mix_spk175_pwgan.wav
paddlespeech tts --am fastspeech2_mix --voc hifigan_csmsc --lang mix --input "我们的声学模型使用了 Fast Speech Two, 声码器使用了 Parallel Wave GAN and Hifi GAN." --spk_id 175 --output mix_spk175.wav
```
- 使用 ONNXRuntime 推理：
```bash
paddlespeech tts --input "你好，欢迎使用百度飞桨深度学习框架！" --output default.wav --use_onnx True
paddlespeech tts --am speedyspeech_csmsc --input "你好，欢迎使用百度飞桨深度学习框架！" --output ss.wav --use_onnx True
paddlespeech tts --voc mb_melgan_csmsc --input "你好，欢迎使用百度飞桨深度学习框架！" --output mb.wav --use_onnx True
paddlespeech tts --voc pwgan_csmsc --input "你好，欢迎使用百度飞桨深度学习框架！" --output pwgan.wav --use_onnx True
paddlespeech tts --am fastspeech2_aishell3 --voc pwgan_aishell3 --input "你好，欢迎使用百度飞桨深度学习框架！" --spk_id 0 --output aishell3_fs2_pwgan.wav --use_onnx True
paddlespeech tts --am fastspeech2_aishell3 --voc hifigan_aishell3 --input "你好，欢迎使用百度飞桨深度学习框架！" --spk_id 0 --output aishell3_fs2_hifigan.wav --use_onnx True
paddlespeech tts --am fastspeech2_ljspeech --voc pwgan_ljspeech --lang en --input "Life was like a box of chocolates, you never know what you're gonna get." --output lj_fs2_pwgan.wav --use_onnx True
paddlespeech tts --am fastspeech2_ljspeech --voc hifigan_ljspeech --lang en --input "Life was like a box of chocolates, you never know what you're gonna get." --output lj_fs2_hifigan.wav --use_onnx True
paddlespeech tts --am fastspeech2_vctk --voc pwgan_vctk --input "Life was like a box of chocolates, you never know what you're gonna get." --lang en --spk_id 0 --output vctk_fs2_pwgan.wav --use_onnx True
paddlespeech tts --am fastspeech2_vctk --voc hifigan_vctk --input "Life was like a box of chocolates, you never know what you're gonna get." --lang en --spk_id 0 --output vctk_fs2_hifigan.wav --use_onnx True
```
使用方法：
......@@ -84,6 +94,8 @@
- `lang`：TTS 任务的语言， 默认值：`zh`。
- `device`：执行预测的设备， 默认值：当前系统下 paddlepaddle 的默认 device。
- `output`：输出音频的路径， 默认值：`output.wav`。
- `use_onnx`: 是否使用 ONNXRuntime 进行推理。
- `fs`: 使用特定 ONNX 模型时的采样率。
输出：
```bash
......@@ -91,31 +103,44 @@
```
- Python API
```python
import paddle
from paddlespeech.cli.tts import TTSExecutor
tts_executor = TTSExecutor()
wav_file = tts_executor(
text='今天的天气不错啊',
output='output.wav',
am='fastspeech2_csmsc',
am_config=None,
am_ckpt=None,
am_stat=None,
spk_id=0,
phones_dict=None,
tones_dict=None,
speaker_dict=None,
voc='pwgan_csmsc',
voc_config=None,
voc_ckpt=None,
voc_stat=None,
lang='zh',
device=paddle.get_device())
print('Wave file has been generated: {}'.format(wav_file))
```
- 动态图推理:
```python
import paddle
from paddlespeech.cli.tts import TTSExecutor
tts_executor = TTSExecutor()
wav_file = tts_executor(
text='今天的天气不错啊',
output='output.wav',
am='fastspeech2_csmsc',
am_config=None,
am_ckpt=None,
am_stat=None,
spk_id=0,
phones_dict=None,
tones_dict=None,
speaker_dict=None,
voc='pwgan_csmsc',
voc_config=None,
voc_ckpt=None,
voc_stat=None,
lang='zh',
device=paddle.get_device())
print('Wave file has been generated: {}'.format(wav_file))
```
- ONNXRuntime 推理:
```python
from paddlespeech.cli.tts import TTSExecutor
tts_executor = TTSExecutor()
wav_file = tts_executor(
text='对数据集进行预处理',
output='output.wav',
am='fastspeech2_csmsc',
voc='hifigan_csmsc',
lang='zh',
use_onnx=True,
cpu_threads=2)
```
输出：
```bash
Wave file has been generated: output.wav
......
docs/requirements.txt
浏览文件 @ 598eb1a5
myst-parser
numpydoc
recommonmark>=0.5.0
sphinx
sphinx-autobuild
sphinx-markdown-tables
sphinx_rtd_theme
paddlepaddle>=2.2.2
braceexpand
colorlog
editdistance
fastapi
g2p_en
g2pM
h5py
......@@ -14,40 +9,45 @@ inflect
jieba
jsonlines
kaldiio
keyboard
librosa==0.8.1
loguru
matplotlib
myst-parser
nara_wpe
numpydoc
onnxruntime==1.10.0
opencc
pandas
paddlenlp
paddlepaddle>=2.2.2
paddlespeech_feat
pandas
pathos == 0.2.8
pattern_singleton
Pillow>=9.0.0
praatio==5.0.0
pypinyin
prettytable
pypinyin<=0.44.0
pypinyin-dict
python-dateutil
pyworld==0.2.12
recommonmark>=0.5.0
resampy==0.2.2
sacrebleu
scipy
sentencepiece~=0.1.96
soundfile~=0.10
sphinx
sphinx-autobuild
sphinx-markdown-tables
sphinx_rtd_theme
textgrid
timer
tqdm
typeguard
uvicorn
visualdl
webrtcvad
websockets
yacs~=0.1.8
prettytable
zhon
colorlog
pathos == 0.2.8
fastapi
websockets
keyboard
uvicorn
pattern_singleton
braceexpand
\ No newline at end of file
docs/source/conf.py
浏览文件 @ 598eb1a5
......@@ -20,10 +20,11 @@
# If extensions (or modules to document with autodoc) are in another directory,
# add these directories to sys.path here. If the directory is relative to the
# documentation root, use os.path.abspath to make it absolute, like shown here.
import os
import sys
import recommonmark.parser
import sphinx_rtd_theme
import sys
import os
sys.path.insert(0, os.path.abspath('../..'))
autodoc_mock_imports = ["soundfile", "librosa"]
......
examples/aishell3/README.md
浏览文件 @ 598eb1a5
......@@ -10,4 +10,5 @@
* voc3 - MultiBand MelGAN
* vc0 - Tacotron2 Voice Cloning with GE2E
* vc1 - FastSpeech2 Voice Cloning with GE2E
* vc2 - FastSpeech2 Voice Cloning with ECAPA-TDNN
* ernie_sat - ERNIE-SAT
examples/aishell3/ernie_sat/run.sh
浏览文件 @ 598eb1a5
......@@ -3,7 +3,7 @@
set -e
source path.sh
gpus=0,1
gpus=0,1,2,3,4,5,6,7
stage=0
stop_stage=100
......
examples/aishell3/tts3/run.sh
浏览文件 @ 598eb1a5
......@@ -44,8 +44,8 @@ fi
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
# install paddle2onnx
version=$(echo `pip list |grep "paddle2onnx"` |awk -F" " '{print $2}')
if [[ -z "$version" || ${version} != '0.9.8' ]]; then
pip install paddle2onnx==0.9.8
if [[ -z "$version" || ${version} != '1.0.0' ]]; then
pip install paddle2onnx==1.0.0
fi
./local/paddle2onnx.sh ${train_output_path} inference inference_onnx fastspeech2_aishell3
# considering the balance between speed and quality, we recommend that you use hifigan as vocoder
......
examples/aishell3/vc1/README.md
浏览文件 @ 598eb1a5
......@@ -99,7 +99,7 @@ CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_p
The synthesizing step is very similar to that one of [tts3](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/aishell3/tts3), but we should set `--voice-cloning=True` when calling `${BIN_DIR}/../synthesize.py`.
### Voice Cloning
Assume there are some reference audios in `./ref_audio`
Assume there are some reference audios in `./ref_audio`
```text
ref_audio
├── 001238.wav
......@@ -116,7 +116,7 @@ CUDA_VISIBLE_DEVICES=${gpus} ./local/voice_cloning.sh ${conf_path} ${train_outpu
Model | Step | eval/loss | eval/l1_loss | eval/duration_loss | eval/pitch_loss| eval/energy_loss
:-------------:| :------------:| :-----: | :-----: | :--------: |:--------:|:---------:
default|2(gpu) x 96400|0.99699|0.62013|0.53057|0.11954| 0.20426|
default|2(gpu) x 96400|0.99699|0.62013|0.053057|0.11954| 0.20426|
FastSpeech2 checkpoint contains files listed below.
(There is no need for `speaker_id_map.txt` here )
......
examples/aishell3/vc2/README.md 0 → 100644
浏览文件 @ 598eb1a5
# FastSpeech2 + AISHELL-3 Voice Cloning (ECAPA-TDNN)
This example contains code used to train a [FastSpeech2](https://arxiv.org/abs/2006.04558) model with [AISHELL-3](http://www.aishelltech.com/aishell_3). The trained model can be used in Voice Cloning Task, We refer to the model structure of [Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis](https://arxiv.org/pdf/1806.04558.pdf). The general steps are as follows:
1. Speaker Encoder: We use Speaker Verification to train a speaker encoder. Datasets used in this task are different from those used in `FastSpeech2` because the transcriptions are not needed, we use more datasets, refer to [ECAPA-TDNN](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/voxceleb/sv0).
2. Synthesizer: We use the trained speaker encoder to generate speaker embedding for each sentence in AISHELL-3. This embedding is an extra input of `FastSpeech2` which will be concated with encoder outputs.
3. Vocoder: We use [Parallel Wave GAN](http://arxiv.org/abs/1910.11480) as the neural Vocoder, refer to [voc1](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/aishell3/voc1).
## Dataset
### Download and Extract
Download AISHELL-3 from it's [Official Website](http://www.aishelltech.com/aishell_3) and extract it to `~/datasets`. Then the dataset is in the directory `~/datasets/data_aishell3`.
### Get MFA Result and Extract
We use [MFA2.x](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) to get durations for aishell3_fastspeech2.
You can download from here [aishell3_alignment_tone.tar.gz](https://paddlespeech.bj.bcebos.com/MFA/AISHELL-3/with_tone/aishell3_alignment_tone.tar.gz), or train your MFA model reference to [mfa example](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/mfa) (use MFA1.x now) of our repo.
## Get Started
Assume the path to the dataset is `~/datasets/data_aishell3`.
Assume the path to the MFA result of AISHELL-3 is `./aishell3_alignment_tone`.
Run the command below to
1. **source path**.
2. preprocess the dataset.
3. train the model.
4. synthesize waveform from `metadata.jsonl`.
5. start a voice cloning inference.
```bash
./run.sh
```
You can choose a range of stages you want to run, or set `stage` equal to `stop-stage` to use only one stage, for example, running the following command will only preprocess the dataset.
```bash
./run.sh --stage 0 --stop-stage 0
```
### Data Preprocessing
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/preprocess.sh ${conf_path}
```
When it is done. A `dump` folder is created in the current directory. The structure of the dump folder is listed below.
```text
dump
├── dev
│ ├── norm
│ └── raw
├── embed
│ ├── SSB0005
│ ├── SSB0009
│ ├── ...
│ └── ...
├── phone_id_map.txt
├── speaker_id_map.txt
├── test
│ ├── norm
│ └── raw
└── train
├── energy_stats.npy
├── norm
├── pitch_stats.npy
├── raw
└── speech_stats.npy
```
The `embed` contains the generated speaker embedding for each sentence in AISHELL-3, which has the same file structure with wav files and the format is `.npy`.
The computing time of utterance embedding can be x hours.
The dataset is split into 3 parts, namely `train`, `dev`, and` test`, each of which contains a `norm` and `raw` subfolder. The raw folder contains speech、pitch and energy features of each utterance, while the norm folder contains normalized ones. The statistics used to normalize features are computed from the training set, which is located in `dump/train/*_stats.npy`.
Also, there is a `metadata.jsonl` in each subfolder. It is a table-like file that contains phones, text_lengths, speech_lengths, durations, the path of speech features, the path of pitch features, the path of energy features, speaker, and id of each utterance.
The preprocessing step is very similar to that one of [tts3](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/aishell3/tts3), but there is one more `ECAPA-TDNN/inference` step here.
### Model Training
`./local/train.sh` calls `${BIN_DIR}/train.py`.
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path}
```
The training step is very similar to that one of [tts3](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/aishell3/tts3), but we should set `--voice-cloning=True` when calling `${BIN_DIR}/train.py`.
### Synthesizing
We use [parallel wavegan](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/aishell3/voc1) as the neural vocoder.
Download pretrained parallel wavegan model from [pwg_aishell3_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/pwgan/pwg_aishell3_ckpt_0.5.zip) and unzip it.
```bash
unzip pwg_aishell3_ckpt_0.5.zip
```
Parallel WaveGAN checkpoint contains files listed below.
```text
pwg_aishell3_ckpt_0.5
├── default.yaml # default config used to train parallel wavegan
├── feats_stats.npy # statistics used to normalize spectrogram when training parallel wavegan
└── snapshot_iter_1000000.pdz # generator parameters of parallel wavegan
```
`./local/synthesize.sh` calls `${BIN_DIR}/../synthesize.py`, which can synthesize waveform from `metadata.jsonl`.
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name}
```
The synthesizing step is very similar to that one of [tts3](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/aishell3/tts3), but we should set `--voice-cloning=True` when calling `${BIN_DIR}/../synthesize.py`.
### Voice Cloning
Assume there are some reference audios in `./ref_audio` (the format must be wav here)
```text
ref_audio
├── 001238.wav
├── LJ015-0254.wav
└── audio_self_test.wav
```
`./local/voice_cloning.sh` calls `${BIN_DIR}/../voice_cloning.py`
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/voice_cloning.sh ${conf_path} ${train_output_path} ${ckpt_name} ${ref_audio_dir}
```
## Pretrained Model
- [fastspeech2_aishell3_ckpt_vc2_1.2.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_aishell3_ckpt_vc2_1.2.0.zip)
Model | Step | eval/loss | eval/l1_loss | eval/duration_loss | eval/pitch_loss| eval/energy_loss
:-------------:| :------------:| :-----: | :-----: | :--------: |:--------:|:---------:
default|2(gpu) x 96400|0.991855|0.599517|0.052142|0.094877| 0.245318|
FastSpeech2 checkpoint contains files listed below.
(There is no need for `speaker_id_map.txt` here )
```text
fastspeech2_aishell3_ckpt_vc2_1.2.0
├── default.yaml # default config used to train fastspeech2
├── energy_stats.npy # statistics used to normalize energy when training fastspeech2
├── phone_id_map.txt # phone vocabulary file when training fastspeech2
├── pitch_stats.npy # statistics used to normalize pitch when training fastspeech2
├── snapshot_iter_96400.pdz # model parameters and optimizer states
└── speech_stats.npy # statistics used to normalize spectrogram when training fastspeech2
```
examples/aishell3/vc2/conf/default.yaml 0 → 100644
浏览文件 @ 598eb1a5
###########################################################
# FEATURE EXTRACTION SETTING #
###########################################################
fs: 24000 # sr
n_fft: 2048 # FFT size (samples).
n_shift: 300 # Hop size (samples). 12.5ms
win_length: 1200 # Window length (samples). 50ms
# If set to null, it will be the same as fft_size.
window: "hann" # Window function.
# Only used for feats_type != raw
fmin: 80 # Minimum frequency of Mel basis.
fmax: 7600 # Maximum frequency of Mel basis.
n_mels: 80 # The number of mel basis.
# Only used for the model using pitch features (e.g. FastSpeech2)
f0min: 80 # Minimum f0 for pitch extraction.
f0max: 400 # Maximum f0 for pitch extraction.
###########################################################
# DATA SETTING #
###########################################################
batch_size: 64
num_workers: 2
###########################################################
# MODEL SETTING #
###########################################################
model:
adim: 384 # attention dimension
aheads: 2 # number of attention heads
elayers: 4 # number of encoder layers
eunits: 1536 # number of encoder ff units
dlayers: 4 # number of decoder layers
dunits: 1536 # number of decoder ff units
positionwise_layer_type: conv1d # type of position-wise layer
positionwise_conv_kernel_size: 3 # kernel size of position wise conv layer
duration_predictor_layers: 2 # number of layers of duration predictor
duration_predictor_chans: 256 # number of channels of duration predictor
duration_predictor_kernel_size: 3 # filter size of duration predictor
postnet_layers: 5 # number of layers of postnset
postnet_filts: 5 # filter size of conv layers in postnet
postnet_chans: 256 # number of channels of conv layers in postnet
use_scaled_pos_enc: True # whether to use scaled positional encoding
encoder_normalize_before: True # whether to perform layer normalization before the input
decoder_normalize_before: True # whether to perform layer normalization before the input
reduction_factor: 1 # reduction factor
init_type: xavier_uniform # initialization type
init_enc_alpha: 1.0 # initial value of alpha of encoder scaled position encoding
init_dec_alpha: 1.0 # initial value of alpha of decoder scaled position encoding
transformer_enc_dropout_rate: 0.2 # dropout rate for transformer encoder layer
transformer_enc_positional_dropout_rate: 0.2 # dropout rate for transformer encoder positional encoding
transformer_enc_attn_dropout_rate: 0.2 # dropout rate for transformer encoder attention layer
transformer_dec_dropout_rate: 0.2 # dropout rate for transformer decoder layer
transformer_dec_positional_dropout_rate: 0.2 # dropout rate for transformer decoder positional encoding
transformer_dec_attn_dropout_rate: 0.2 # dropout rate for transformer decoder attention layer
pitch_predictor_layers: 5 # number of conv layers in pitch predictor
pitch_predictor_chans: 256 # number of channels of conv layers in pitch predictor
pitch_predictor_kernel_size: 5 # kernel size of conv leyers in pitch predictor
pitch_predictor_dropout: 0.5 # dropout rate in pitch predictor
pitch_embed_kernel_size: 1 # kernel size of conv embedding layer for pitch
pitch_embed_dropout: 0.0 # dropout rate after conv embedding layer for pitch
stop_gradient_from_pitch_predictor: True # whether to stop the gradient from pitch predictor to encoder
energy_predictor_layers: 2 # number of conv layers in energy predictor
energy_predictor_chans: 256 # number of channels of conv layers in energy predictor
energy_predictor_kernel_size: 3 # kernel size of conv leyers in energy predictor
energy_predictor_dropout: 0.5 # dropout rate in energy predictor
energy_embed_kernel_size: 1 # kernel size of conv embedding layer for energy
energy_embed_dropout: 0.0 # dropout rate after conv embedding layer for energy
stop_gradient_from_energy_predictor: False # whether to stop the gradient from energy predictor to encoder
spk_embed_dim: 192 # speaker embedding dimension
spk_embed_integration_type: concat # speaker embedding integration type
###########################################################
# UPDATER SETTING #
###########################################################
updater:
use_masking: True # whether to apply masking for padded part in loss calculation
###########################################################
# OPTIMIZER SETTING #
###########################################################
optimizer:
optim: adam # optimizer type
learning_rate: 0.001 # learning rate
###########################################################
# TRAINING SETTING #
###########################################################
max_epoch: 200
num_snapshots: 5
###########################################################
# OTHER SETTING #
###########################################################
seed: 10086
examples/aishell3/vc2/local/preprocess.sh 0 → 100755
浏览文件 @ 598eb1a5
#!/bin/bash
stage=0
stop_stage=100
config_path=$1
# gen speaker embedding
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
python3 ${BIN_DIR}/vc2_infer.py \
--input=~/datasets/data_aishell3/train/wav/ \
--output=dump/embed \
--num-cpu=20
fi
# copy from tts3/preprocess
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# get durations from MFA's result
echo "Generate durations.txt from MFA results ..."
python3 ${MAIN_ROOT}/utils/gen_duration_from_textgrid.py \
--inputdir=./aishell3_alignment_tone \
--output durations.txt \
--config=${config_path}
fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
# extract features
echo "Extract features ..."
python3 ${BIN_DIR}/preprocess.py \
--dataset=aishell3 \
--rootdir=~/datasets/data_aishell3/ \
--dumpdir=dump \
--dur-file=durations.txt \
--config=${config_path} \
--num-cpu=20 \
--cut-sil=True \
--spk_emb_dir=dump/embed
fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# get features' stats(mean and std)
echo "Get features' stats ..."
python3 ${MAIN_ROOT}/utils/compute_statistics.py \
--metadata=dump/train/raw/metadata.jsonl \
--field-name="speech"
python3 ${MAIN_ROOT}/utils/compute_statistics.py \
--metadata=dump/train/raw/metadata.jsonl \
--field-name="pitch"
python3 ${MAIN_ROOT}/utils/compute_statistics.py \
--metadata=dump/train/raw/metadata.jsonl \
--field-name="energy"
fi
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
# normalize and covert phone/speaker to id, dev and test should use train's stats
echo "Normalize ..."
python3 ${BIN_DIR}/normalize.py \
--metadata=dump/train/raw/metadata.jsonl \
--dumpdir=dump/train/norm \
--speech-stats=dump/train/speech_stats.npy \
--pitch-stats=dump/train/pitch_stats.npy \
--energy-stats=dump/train/energy_stats.npy \
--phones-dict=dump/phone_id_map.txt \
--speaker-dict=dump/speaker_id_map.txt
python3 ${BIN_DIR}/normalize.py \
--metadata=dump/dev/raw/metadata.jsonl \
--dumpdir=dump/dev/norm \
--speech-stats=dump/train/speech_stats.npy \
--pitch-stats=dump/train/pitch_stats.npy \
--energy-stats=dump/train/energy_stats.npy \
--phones-dict=dump/phone_id_map.txt \
--speaker-dict=dump/speaker_id_map.txt
python3 ${BIN_DIR}/normalize.py \
--metadata=dump/test/raw/metadata.jsonl \
--dumpdir=dump/test/norm \
--speech-stats=dump/train/speech_stats.npy \
--pitch-stats=dump/train/pitch_stats.npy \
--energy-stats=dump/train/energy_stats.npy \
--phones-dict=dump/phone_id_map.txt \
--speaker-dict=dump/speaker_id_map.txt
fi
examples/aishell3/vc2/local/synthesize.sh 0 → 100755
浏览文件 @ 598eb1a5
#!/bin/bash
config_path=$1
train_output_path=$2
ckpt_name=$3
FLAGS_allocator_strategy=naive_best_fit \
FLAGS_fraction_of_gpu_memory_to_use=0.01 \
python3 ${BIN_DIR}/../synthesize.py \
--am=fastspeech2_aishell3 \
--am_config=${config_path} \
--am_ckpt=${train_output_path}/checkpoints/${ckpt_name} \
--am_stat=dump/train/speech_stats.npy \
--voc=pwgan_aishell3 \
--voc_config=pwg_aishell3_ckpt_0.5/default.yaml \
--voc_ckpt=pwg_aishell3_ckpt_0.5/snapshot_iter_1000000.pdz \
--voc_stat=pwg_aishell3_ckpt_0.5/feats_stats.npy \
--test_metadata=dump/test/norm/metadata.jsonl \
--output_dir=${train_output_path}/test \
--phones_dict=dump/phone_id_map.txt \
--speaker_dict=dump/speaker_id_map.txt \
--voice-cloning=True
examples/aishell3/vc2/local/train.sh 0 → 100755
浏览文件 @ 598eb1a5
#!/bin/bash
config_path=$1
train_output_path=$2
python3 ${BIN_DIR}/train.py \
--train-metadata=dump/train/norm/metadata.jsonl \
--dev-metadata=dump/dev/norm/metadata.jsonl \
--config=${config_path} \
--output-dir=${train_output_path} \
--ngpu=2 \
--phones-dict=dump/phone_id_map.txt \
--voice-cloning=True
\ No newline at end of file
examples/aishell3/vc2/local/voice_cloning.sh 0 → 100755
浏览文件 @ 598eb1a5
#!/bin/bash
config_path=$1
train_output_path=$2
ckpt_name=$3
ref_audio_dir=$4
FLAGS_allocator_strategy=naive_best_fit \
FLAGS_fraction_of_gpu_memory_to_use=0.01 \
python3 ${BIN_DIR}/../voice_cloning.py \
--am=fastspeech2_aishell3 \
--am_config=${config_path} \
--am_ckpt=${train_output_path}/checkpoints/${ckpt_name} \
--am_stat=dump/train/speech_stats.npy \
--voc=pwgan_aishell3 \
--voc_config=pwg_aishell3_ckpt_0.5/default.yaml \
--voc_ckpt=pwg_aishell3_ckpt_0.5/snapshot_iter_1000000.pdz \
--voc_stat=pwg_aishell3_ckpt_0.5/feats_stats.npy \
--text="凯莫瑞安联合体的经济崩溃迫在眉睫。" \
--input-dir=${ref_audio_dir} \
--output-dir=${train_output_path}/vc_syn \
--phones-dict=dump/phone_id_map.txt \
--use_ecapa=True
examples/aishell3/vc2/path.sh 0 → 100755
浏览文件 @ 598eb1a5
#!/bin/bash
export MAIN_ROOT=`realpath ${PWD}/../../../`
export PATH=${MAIN_ROOT}:${MAIN_ROOT}/utils:${PATH}
export LC_ALL=C
export PYTHONDONTWRITEBYTECODE=1
# Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
export PYTHONIOENCODING=UTF-8
export PYTHONPATH=${MAIN_ROOT}:${PYTHONPATH}
MODEL=fastspeech2
export BIN_DIR=${MAIN_ROOT}/paddlespeech/t2s/exps/${MODEL}
examples/aishell3/vc2/run.sh 0 → 100755
浏览文件 @ 598eb1a5
#!/bin/bash
set -e
source path.sh
gpus=0,1
stage=0
stop_stage=100
conf_path=conf/default.yaml
train_output_path=exp/default
ckpt_name=snapshot_iter_96400.pdz
ref_audio_dir=ref_audio
# with the following command, you can choose the stage range you want to run
# such as `./run.sh --stage 0 --stop-stage 0`
# this can not be mixed use with `$1`, `$2` ...
source ${MAIN_ROOT}/utils/parse_options.sh || exit 1
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
# prepare data
CUDA_VISIBLE_DEVICES=${gpus} ./local/preprocess.sh ${conf_path} || exit -1
fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# train model, all `ckpt` under `train_output_path/checkpoints/` dir
CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path} || exit -1
fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
# synthesize, vocoder is pwgan
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1
fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# synthesize, vocoder is pwgan
CUDA_VISIBLE_DEVICES=${gpus} ./local/voice_cloning.sh ${conf_path} ${train_output_path} ${ckpt_name} ${ref_audio_dir} || exit -1
fi
examples/aishell3/vits-vc/README.md 0 → 100644
浏览文件 @ 598eb1a5
# VITS with AISHELL-3
This example contains code used to train a [VITS](https://arxiv.org/abs/2106.06103) model with [AISHELL-3](http://www.aishelltech.com/aishell_3). The trained model can be used in Voice Cloning Task, We refer to the model structure of [Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis](https://arxiv.org/pdf/1806.04558.pdf). The general steps are as follows:
1. Speaker Encoder: We use Speaker Verification to train a speaker encoder. Datasets used in this task are different from those used in `VITS` because the transcriptions are not needed, we use more datasets, refer to [ge2e](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/ge2e).
2. Synthesizer and Vocoder: We use the trained speaker encoder to generate speaker embedding for each sentence in AISHELL-3. This embedding is an extra input of `VITS` which will be concated with encoder outputs. The vocoder is part of `VITS` due to its special structure.
## Dataset
### Download and Extract
Download AISHELL-3 from it's [Official Website](http://www.aishelltech.com/aishell_3) and extract it to `~/datasets`. Then the dataset is in the directory `~/datasets/data_aishell3`.
### Get MFA Result and Extract
We use [MFA2.x](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) to get phonemes for VITS, the durations of MFA are not needed here.
You can download from here [aishell3_alignment_tone.tar.gz](https://paddlespeech.bj.bcebos.com/MFA/AISHELL-3/with_tone/aishell3_alignment_tone.tar.gz), or train your MFA model reference to [mfa example](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/mfa) (use MFA1.x now) of our repo.
## Pretrained GE2E Model
We use pretrained GE2E model to generate speaker embedding for each sentence.
Download pretrained GE2E model from here [ge2e_ckpt_0.3.zip](https://bj.bcebos.com/paddlespeech/Parakeet/released_models/ge2e/ge2e_ckpt_0.3.zip), and `unzip` it.
## Get Started
Assume the path to the dataset is `~/datasets/data_aishell3`.
Assume the path to the MFA result of AISHELL-3 is `./aishell3_alignment_tone`.
Assume the path to the pretrained ge2e model is `./ge2e_ckpt_0.3`.
Run the command below to
1. **source path**.
2. preprocess the dataset.
3. train the model.
4. synthesize waveform from `metadata.jsonl`.
5. start a voice cloning inference.
```bash
./run.sh
```
You can choose a range of stages you want to run, or set `stage` equal to `stop-stage` to use only one stage, for example, running the following command will only preprocess the dataset.
```bash
./run.sh --stage 0 --stop-stage 0
```
### Data Preprocessing
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/preprocess.sh ${conf_path} ${ge2e_ckpt_path}
```
When it is done. A `dump` folder is created in the current directory. The structure of the dump folder is listed below.
```text
dump
├── dev
│   ├── norm
│   └── raw
├── embed
│ ├── SSB0005
│ ├── SSB0009
│ ├── ...
│ └── ...
├── phone_id_map.txt
├── speaker_id_map.txt
├── test
│   ├── norm
│   └── raw
└── train
├── feats_stats.npy
├── norm
└── raw
```
The `embed` contains the generated speaker embedding for each sentence in AISHELL-3, which has the same file structure with wav files and the format is `.npy`.
The computing time of utterance embedding can be x hours.
The dataset is split into 3 parts, namely `train`, `dev`, and` test`, each of which contains a `norm` and `raw` subfolder. The raw folder contains wave and linear spectrogram of each utterance, while the norm folder contains normalized ones. The statistics used to normalize features are computed from the training set, which is located in `dump/train/feats_stats.npy`.
Also, there is a `metadata.jsonl` in each subfolder. It is a table-like file that contains phones, text_lengths, feats, feats_lengths, the path of linear spectrogram features, the path of raw waves, speaker, and the id of each utterance.
The preprocessing step is very similar to that one of [vits](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/aishell3/vits), but there is one more `ge2e/inference` step here.
### Model Training
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path}
```
The training step is very similar to that one of [vits](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/aishell3/vits), but we should set `--voice-cloning=True` when calling `${BIN_DIR}/train.py`.
### Synthesizing
`./local/synthesize.sh` calls `${BIN_DIR}/synthesize.py`, which can synthesize waveform from `metadata.jsonl`.
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name}
```
```text
usage: synthesize.py [-h] [--config CONFIG] [--ckpt CKPT]
[--phones_dict PHONES_DICT] [--speaker_dict SPEAKER_DICT]
[--voice-cloning VOICE_CLONING] [--ngpu NGPU]
[--test_metadata TEST_METADATA] [--output_dir OUTPUT_DIR]
Synthesize with VITS
optional arguments:
-h, --help show this help message and exit
--config CONFIG Config of VITS.
--ckpt CKPT Checkpoint file of VITS.
--phones_dict PHONES_DICT
phone vocabulary file.
--speaker_dict SPEAKER_DICT
speaker id map file.
--voice-cloning VOICE_CLONING
whether training voice cloning model.
--ngpu NGPU if ngpu == 0, use cpu.
--test_metadata TEST_METADATA
test metadata.
--output_dir OUTPUT_DIR
output dir.
```
The synthesizing step is very similar to that one of [vits](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/aishell3/vits), but we should set `--voice-cloning=True` when calling `${BIN_DIR}/../synthesize.py`.
### Voice Cloning
Assume there are some reference audios in `./ref_audio`
```text
ref_audio
├── 001238.wav
├── LJ015-0254.wav
└── audio_self_test.mp3
```
`./local/voice_cloning.sh` calls `${BIN_DIR}/voice_cloning.py`
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/voice_cloning.sh ${conf_path} ${train_output_path} ${ckpt_name} ${ge2e_params_path} ${add_blank} ${ref_audio_dir}
```
If you want to convert a speaker audio file to refered speaker, run:
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/voice_cloning.sh ${conf_path} ${train_output_path} ${ckpt_name} ${ge2e_params_path} ${add_blank} ${ref_audio_dir} ${src_audio_path}
```
<!-- TODO display these after we trained the model -->
<!--
## Pretrained Model
The pretrained model can be downloaded here:
- [vits_vc_aishell3_ckpt_1.1.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/vits/vits_vc_aishell3_ckpt_1.1.0.zip) (add_blank=true)
VITS checkpoint contains files listed below.
(There is no need for `speaker_id_map.txt` here )
```text
vits_vc_aishell3_ckpt_1.1.0
├── default.yaml # default config used to train vitx
├── phone_id_map.txt # phone vocabulary file when training vits
└── snapshot_iter_333000.pdz # model parameters and optimizer states
```
ps: This ckpt is not good enough, a better result is training
-->
examples/aishell3/vits-vc/conf/default.yaml 0 → 100644
浏览文件 @ 598eb1a5
# This configuration tested on 4 GPUs (V100) with 32GB GPU
# memory. It takes around 2 weeks to finish the training
# but 100k iters model should generate reasonable results.
###########################################################
# FEATURE EXTRACTION SETTING #
###########################################################
fs: 22050 # sr
n_fft: 1024 # FFT size (samples).
n_shift: 256 # Hop size (samples). 12.5ms
win_length: null # Window length (samples). 50ms
# If set to null, it will be the same as fft_size.
window: "hann" # Window function.
##########################################################
# TTS MODEL SETTING #
##########################################################
model:
# generator related
generator_type: vits_generator
generator_params:
hidden_channels: 192
spk_embed_dim: 256
global_channels: 256
segment_size: 32
text_encoder_attention_heads: 2
text_encoder_ffn_expand: 4
text_encoder_blocks: 6
text_encoder_positionwise_layer_type: "conv1d"
text_encoder_positionwise_conv_kernel_size: 3
text_encoder_positional_encoding_layer_type: "rel_pos"
text_encoder_self_attention_layer_type: "rel_selfattn"
text_encoder_activation_type: "swish"
text_encoder_normalize_before: True
text_encoder_dropout_rate: 0.1
text_encoder_positional_dropout_rate: 0.0
text_encoder_attention_dropout_rate: 0.1
use_macaron_style_in_text_encoder: True
use_conformer_conv_in_text_encoder: False
text_encoder_conformer_kernel_size: -1
decoder_kernel_size: 7
decoder_channels: 512
decoder_upsample_scales: [8, 8, 2, 2]
decoder_upsample_kernel_sizes: [16, 16, 4, 4]
decoder_resblock_kernel_sizes: [3, 7, 11]
decoder_resblock_dilations: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
use_weight_norm_in_decoder: True
posterior_encoder_kernel_size: 5
posterior_encoder_layers: 16
posterior_encoder_stacks: 1
posterior_encoder_base_dilation: 1
posterior_encoder_dropout_rate: 0.0
use_weight_norm_in_posterior_encoder: True
flow_flows: 4
flow_kernel_size: 5
flow_base_dilation: 1
flow_layers: 4
flow_dropout_rate: 0.0
use_weight_norm_in_flow: True
use_only_mean_in_flow: True
stochastic_duration_predictor_kernel_size: 3
stochastic_duration_predictor_dropout_rate: 0.5
stochastic_duration_predictor_flows: 4
stochastic_duration_predictor_dds_conv_layers: 3
# discriminator related
discriminator_type: hifigan_multi_scale_multi_period_discriminator
discriminator_params:
scales: 1
scale_downsample_pooling: "AvgPool1D"
scale_downsample_pooling_params:
kernel_size: 4
stride: 2
padding: 2
scale_discriminator_params:
in_channels: 1
out_channels: 1
kernel_sizes: [15, 41, 5, 3]
channels: 128
max_downsample_channels: 1024
max_groups: 16
bias: True
downsample_scales: [2, 2, 4, 4, 1]
nonlinear_activation: "leakyrelu"
nonlinear_activation_params:
negative_slope: 0.1
use_weight_norm: True
use_spectral_norm: False
follow_official_norm: False
periods: [2, 3, 5, 7, 11]
period_discriminator_params:
in_channels: 1
out_channels: 1
kernel_sizes: [5, 3]
channels: 32
downsample_scales: [3, 3, 3, 3, 1]
max_downsample_channels: 1024
bias: True
nonlinear_activation: "leakyrelu"
nonlinear_activation_params:
negative_slope: 0.1
use_weight_norm: True
use_spectral_norm: False
# others
sampling_rate: 22050 # needed in the inference for saving wav
cache_generator_outputs: True # whether to cache generator outputs in the training
###########################################################
# LOSS SETTING #
###########################################################
# loss function related
generator_adv_loss_params:
average_by_discriminators: False # whether to average loss value by #discriminators
loss_type: mse # loss type, "mse" or "hinge"
discriminator_adv_loss_params:
average_by_discriminators: False # whether to average loss value by #discriminators
loss_type: mse # loss type, "mse" or "hinge"
feat_match_loss_params:
average_by_discriminators: False # whether to average loss value by #discriminators
average_by_layers: False # whether to average loss value by #layers of each discriminator
include_final_outputs: True # whether to include final outputs for loss calculation
mel_loss_params:
fs: 22050 # must be the same as the training data
fft_size: 1024 # fft points
hop_size: 256 # hop size
win_length: null # window length
window: hann # window type
num_mels: 80 # number of Mel basis
fmin: 0 # minimum frequency for Mel basis
fmax: null # maximum frequency for Mel basis
log_base: null # null represent natural log
###########################################################
# ADVERSARIAL LOSS SETTING #
###########################################################
lambda_adv: 1.0 # loss scaling coefficient for adversarial loss
lambda_mel: 45.0 # loss scaling coefficient for Mel loss
lambda_feat_match: 2.0 # loss scaling coefficient for feat match loss
lambda_dur: 1.0 # loss scaling coefficient for duration loss
lambda_kl: 1.0 # loss scaling coefficient for KL divergence loss
# others
sampling_rate: 22050 # needed in the inference for saving wav
cache_generator_outputs: True # whether to cache generator outputs in the training
###########################################################
# DATA LOADER SETTING #
###########################################################
batch_size: 50 # Batch size.
num_workers: 4 # Number of workers in DataLoader.
##########################################################
# OPTIMIZER & SCHEDULER SETTING #
##########################################################
# optimizer setting for generator
generator_optimizer_params:
beta1: 0.8
beta2: 0.99
epsilon: 1.0e-9
weight_decay: 0.0
generator_scheduler: exponential_decay
generator_scheduler_params:
learning_rate: 2.0e-4
gamma: 0.999875
# optimizer setting for discriminator
discriminator_optimizer_params:
beta1: 0.8
beta2: 0.99
epsilon: 1.0e-9
weight_decay: 0.0
discriminator_scheduler: exponential_decay
discriminator_scheduler_params:
learning_rate: 2.0e-4
gamma: 0.999875
generator_first: False # whether to start updating generator first
##########################################################
# OTHER TRAINING SETTING #
##########################################################
num_snapshots: 10 # max number of snapshots to keep while training
train_max_steps: 350000 # Number of training steps. == total_iters / ngpus, total_iters = 1000000
save_interval_steps: 1000 # Interval steps to save checkpoint.
eval_interval_steps: 250 # Interval steps to evaluate the network.
seed: 777 # random seed number
examples/aishell3/vits-vc/local/preprocess.sh 0 → 100755
浏览文件 @ 598eb1a5
#!/bin/bash
stage=0
stop_stage=100
config_path=$1
add_blank=$2
ge2e_ckpt_path=$3
# gen speaker embedding
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
python3 ${MAIN_ROOT}/paddlespeech/vector/exps/ge2e/inference.py \
--input=~/datasets/data_aishell3/train/wav/ \
--output=dump/embed \
--checkpoint_path=${ge2e_ckpt_path}
fi
# copy from tts3/preprocess
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# get durations from MFA's result
echo "Generate durations.txt from MFA results ..."
python3 ${MAIN_ROOT}/utils/gen_duration_from_textgrid.py \
--inputdir=./aishell3_alignment_tone \
--output durations.txt \
--config=${config_path}
fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
# extract features
echo "Extract features ..."
python3 ${BIN_DIR}/preprocess.py \
--dataset=aishell3 \
--rootdir=~/datasets/data_aishell3/ \
--dumpdir=dump \
--dur-file=durations.txt \
--config=${config_path} \
--num-cpu=20 \
--cut-sil=True \
--spk_emb_dir=dump/embed
fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# get features' stats(mean and std)
echo "Get features' stats ..."
python3 ${MAIN_ROOT}/utils/compute_statistics.py \
--metadata=dump/train/raw/metadata.jsonl \
--field-name="feats"
fi
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
# normalize and covert phone/speaker to id, dev and test should use train's stats
echo "Normalize ..."
python3 ${BIN_DIR}/normalize.py \
--metadata=dump/train/raw/metadata.jsonl \
--dumpdir=dump/train/norm \
--feats-stats=dump/train/feats_stats.npy \
--phones-dict=dump/phone_id_map.txt \
--speaker-dict=dump/speaker_id_map.txt \
--add-blank=${add_blank} \
--skip-wav-copy
python3 ${BIN_DIR}/normalize.py \
--metadata=dump/dev/raw/metadata.jsonl \
--dumpdir=dump/dev/norm \
--feats-stats=dump/train/feats_stats.npy \
--phones-dict=dump/phone_id_map.txt \
--speaker-dict=dump/speaker_id_map.txt \
--add-blank=${add_blank} \
--skip-wav-copy
python3 ${BIN_DIR}/normalize.py \
--metadata=dump/test/raw/metadata.jsonl \
--dumpdir=dump/test/norm \
--feats-stats=dump/train/feats_stats.npy \
--phones-dict=dump/phone_id_map.txt \
--speaker-dict=dump/speaker_id_map.txt \
--add-blank=${add_blank} \
--skip-wav-copy
fi
examples/aishell3/vits-vc/local/synthesize.sh 0 → 100755
浏览文件 @ 598eb1a5
#!/bin/bash
config_path=$1
train_output_path=$2
ckpt_name=$3
stage=0
stop_stage=0
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
FLAGS_allocator_strategy=naive_best_fit \
FLAGS_fraction_of_gpu_memory_to_use=0.01 \
python3 ${BIN_DIR}/synthesize.py \
--config=${config_path} \
--ckpt=${train_output_path}/checkpoints/${ckpt_name} \
--phones_dict=dump/phone_id_map.txt \
--test_metadata=dump/test/norm/metadata.jsonl \
--output_dir=${train_output_path}/test \
--voice-cloning=True
fi
examples/aishell3/vits-vc/local/train.sh 0 → 100755
浏览文件 @ 598eb1a5
#!/bin/bash
config_path=$1
train_output_path=$2
# install monotonic_align
cd ${MAIN_ROOT}/paddlespeech/t2s/models/vits/monotonic_align
python3 setup.py build_ext --inplace
cd -
python3 ${BIN_DIR}/train.py \
--train-metadata=dump/train/norm/metadata.jsonl \
--dev-metadata=dump/dev/norm/metadata.jsonl \
--config=${config_path} \
--output-dir=${train_output_path} \
--ngpu=4 \
--phones-dict=dump/phone_id_map.txt \
--voice-cloning=True
examples/aishell3/vits-vc/local/voice_cloning.sh 0 → 100755
浏览文件 @ 598eb1a5
#!/bin/bash
config_path=$1
train_output_path=$2
ckpt_name=$3
ge2e_params_path=$4
add_blank=$5
ref_audio_dir=$6
src_audio_path=$7
FLAGS_allocator_strategy=naive_best_fit \
FLAGS_fraction_of_gpu_memory_to_use=0.01 \
python3 ${BIN_DIR}/voice_cloning.py \
--config=${config_path} \
--ckpt=${train_output_path}/checkpoints/${ckpt_name} \
--ge2e_params_path=${ge2e_params_path} \
--phones_dict=dump/phone_id_map.txt \
--text="凯莫瑞安联合体的经济崩溃迫在眉睫。" \
--audio-path=${src_audio_path} \
--input-dir=${ref_audio_dir} \
--output-dir=${train_output_path}/vc_syn \
--add-blank=${add_blank}
examples/aishell3/vits-vc/path.sh 0 → 100755
浏览文件 @ 598eb1a5
#!/bin/bash
export MAIN_ROOT=`realpath ${PWD}/../../../`
export PATH=${MAIN_ROOT}:${MAIN_ROOT}/utils:${PATH}
export LC_ALL=C
export PYTHONDONTWRITEBYTECODE=1
# Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
export PYTHONIOENCODING=UTF-8
export PYTHONPATH=${MAIN_ROOT}:${PYTHONPATH}
MODEL=vits
export BIN_DIR=${MAIN_ROOT}/paddlespeech/t2s/exps/${MODEL}
\ No newline at end of file
examples/aishell3/vits-vc/run.sh 0 → 100755
浏览文件 @ 598eb1a5
#!/bin/bash
set -e
source path.sh
gpus=0,1,2,3
stage=0
stop_stage=100
conf_path=conf/default.yaml
train_output_path=exp/default
ckpt_name=snapshot_iter_153.pdz
add_blank=true
ref_audio_dir=ref_audio
src_audio_path=''
# not include ".pdparams" here
ge2e_ckpt_path=./ge2e_ckpt_0.3/step-3000000
# include ".pdparams" here
ge2e_params_path=${ge2e_ckpt_path}.pdparams
# with the following command, you can choose the stage range you want to run
# such as `./run.sh --stage 0 --stop-stage 0`
# this can not be mixed use with `$1`, `$2` ...
source ${MAIN_ROOT}/utils/parse_options.sh || exit 1
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
# prepare data
CUDA_VISIBLE_DEVICES=${gpus} ./local/preprocess.sh ${conf_path} ${add_blank} ${ge2e_ckpt_path} || exit -1
fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# train model, all `ckpt` under `train_output_path/checkpoints/` dir
CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path} || exit -1
fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1
fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
CUDA_VISIBLE_DEVICES=${gpus} ./local/voice_cloning.sh ${conf_path} ${train_output_path} ${ckpt_name} \
${ge2e_params_path} ${add_blank} ${ref_audio_dir} ${src_audio_path} || exit -1
fi
examples/aishell3/vits/README.md 0 → 100644
浏览文件 @ 598eb1a5
# VITS with AISHELL-3
This example contains code used to train a [VITS](https://arxiv.org/abs/2106.06103) model with [AISHELL-3](http://www.aishelltech.com/aishell_3).
AISHELL-3 is a large-scale and high-fidelity multi-speaker Mandarin speech corpus that could be used to train multi-speaker Text-to-Speech (TTS) systems.
We use AISHELL-3 to train a multi-speaker VITS model here.
## Dataset
### Download and Extract
Download AISHELL-3 from it's [Official Website](http://www.aishelltech.com/aishell_3) and extract it to `~/datasets`. Then the dataset is in the directory `~/datasets/data_aishell3`.
### Get MFA Result and Extract
We use [MFA2.x](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) to get phonemes for VITS, the durations of MFA are not needed here.
You can download from here [aishell3_alignment_tone.tar.gz](https://paddlespeech.bj.bcebos.com/MFA/AISHELL-3/with_tone/aishell3_alignment_tone.tar.gz), or train your MFA model reference to [mfa example](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/mfa) (use MFA1.x now) of our repo.
## Get Started
Assume the path to the dataset is `~/datasets/data_aishell3`.
Assume the path to the MFA result of AISHELL-3 is `./aishell3_alignment_tone`.
Run the command below to
1. **source path**.
2. preprocess the dataset.
3. train the model.
4. synthesize wavs.
- synthesize waveform from `metadata.jsonl`.
- synthesize waveform from a text file.
```bash
./run.sh
```
You can choose a range of stages you want to run, or set `stage` equal to `stop-stage` to use only one stage, for example, running the following command will only preprocess the dataset.
```bash
./run.sh --stage 0 --stop-stage 0
```
### Data Preprocessing
```bash
./local/preprocess.sh ${conf_path}
```
When it is done. A `dump` folder is created in the current directory. The structure of the dump folder is listed below.
```text
dump
├── dev
│   ├── norm
│   └── raw
├── phone_id_map.txt
├── speaker_id_map.txt
├── test
│   ├── norm
│   └── raw
└── train
├── feats_stats.npy
├── norm
└── raw
```
The dataset is split into 3 parts, namely `train`, `dev`, and` test`, each of which contains a `norm` and `raw` subfolder. The raw folder contains wave and linear spectrogram of each utterance, while the norm folder contains normalized ones. The statistics used to normalize features are computed from the training set, which is located in `dump/train/feats_stats.npy`.
Also, there is a `metadata.jsonl` in each subfolder. It is a table-like file that contains phones, text_lengths, feats, feats_lengths, the path of linear spectrogram features, the path of raw waves, speaker, and the id of each utterance.
### Model Training
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path}
```
`./local/train.sh` calls `${BIN_DIR}/train.py`.
Here's the complete help message.
```text
usage: train.py [-h] [--config CONFIG] [--train-metadata TRAIN_METADATA]
[--dev-metadata DEV_METADATA] [--output-dir OUTPUT_DIR]
[--ngpu NGPU] [--phones-dict PHONES_DICT]
[--speaker-dict SPEAKER_DICT] [--voice-cloning VOICE_CLONING]
Train a VITS model.
optional arguments:
-h, --help show this help message and exit
--config CONFIG config file to overwrite default config.
--train-metadata TRAIN_METADATA
training data.
--dev-metadata DEV_METADATA
dev data.
--output-dir OUTPUT_DIR
output dir.
--ngpu NGPU if ngpu == 0, use cpu.
--phones-dict PHONES_DICT
phone vocabulary file.
--speaker-dict SPEAKER_DICT
speaker id map file for multiple speaker model.
--voice-cloning VOICE_CLONING
whether training voice cloning model.
```
1. `--config` is a config file in yaml format to overwrite the default config, which can be found at `conf/default.yaml`.
2. `--train-metadata` and `--dev-metadata` should be the metadata file in the normalized subfolder of `train` and `dev` in the `dump` folder.
3. `--output-dir` is the directory to save the results of the experiment. Checkpoints are saved in `checkpoints/` inside this directory.
4. `--ngpu` is the number of gpus to use, if ngpu == 0, use cpu.
5. `--phones-dict` is the path of the phone vocabulary file.
6. `--speaker-dict` is the path of the speaker id map file when training a multi-speaker VITS.
### Synthesizing
`./local/synthesize.sh` calls `${BIN_DIR}/synthesize.py`, which can synthesize waveform from `metadata.jsonl`.
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name}
```
```text
usage: synthesize.py [-h] [--config CONFIG] [--ckpt CKPT]
[--phones_dict PHONES_DICT] [--speaker_dict SPEAKER_DICT]
[--voice-cloning VOICE_CLONING] [--ngpu NGPU]
[--test_metadata TEST_METADATA] [--output_dir OUTPUT_DIR]
Synthesize with VITS
optional arguments:
-h, --help show this help message and exit
--config CONFIG Config of VITS.
--ckpt CKPT Checkpoint file of VITS.
--phones_dict PHONES_DICT
phone vocabulary file.
--speaker_dict SPEAKER_DICT
speaker id map file.
--voice-cloning VOICE_CLONING
whether training voice cloning model.
--ngpu NGPU if ngpu == 0, use cpu.
--test_metadata TEST_METADATA
test metadata.
--output_dir OUTPUT_DIR
output dir.
```
`./local/synthesize_e2e.sh` calls `${BIN_DIR}/synthesize_e2e.py`, which can synthesize waveform from text file.
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize_e2e.sh ${conf_path} ${train_output_path} ${ckpt_name}
```
```text
usage: synthesize_e2e.py [-h] [--config CONFIG] [--ckpt CKPT]
[--phones_dict PHONES_DICT]
[--speaker_dict SPEAKER_DICT] [--spk_id SPK_ID]
[--lang LANG]
[--inference_dir INFERENCE_DIR] [--ngpu NGPU]
[--text TEXT] [--output_dir OUTPUT_DIR]
Synthesize with VITS
optional arguments:
-h, --help show this help message and exit
--config CONFIG Config of VITS.
--ckpt CKPT Checkpoint file of VITS.
--phones_dict PHONES_DICT
phone vocabulary file.
--speaker_dict SPEAKER_DICT
speaker id map file.
--spk_id SPK_ID spk id for multi speaker acoustic model
--lang LANG Choose model language. zh or en
--inference_dir INFERENCE_DIR
dir to save inference models
--ngpu NGPU if ngpu == 0, use cpu.
--text TEXT text to synthesize, a 'utt_id sentence' pair per line.
--output_dir OUTPUT_DIR
output dir.
```
1. `--config`, `--ckpt`, `--phones_dict` and `--speaker_dict` are arguments for acoustic model, which correspond to the 3 files in the VITS pretrained model.
2. `--lang` is the model language, which can be `zh` or `en`.
3. `--test_metadata` should be the metadata file in the normalized subfolder of `test` in the `dump` folder.
4. `--text` is the text file, which contains sentences to synthesize.
5. `--output_dir` is the directory to save synthesized audio files.
6. `--ngpu` is the number of gpus to use, if ngpu == 0, use cpu.
<!-- TODO display these after we trained the model -->
<!--
## Pretrained Model
The pretrained model can be downloaded here:
- [vits_aishell3_ckpt_1.1.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/vits/vits_aishell3_ckpt_1.1.0.zip) (add_blank=true)
VITS checkpoint contains files listed below.
```text
vits_aishell3_ckpt_1.1.0
├── default.yaml # default config used to train vitx
├── phone_id_map.txt # phone vocabulary file when training vits
├── speaker_id_map.txt # speaker id map file when training a multi-speaker vits
└── snapshot_iter_333000.pdz # model parameters and optimizer states
```
ps: This ckpt is not good enough, a better result is training
You can use the following scripts to synthesize for `${BIN_DIR}/../sentences.txt` using pretrained VITS.
```bash
source path.sh
add_blank=true
FLAGS_allocator_strategy=naive_best_fit \
FLAGS_fraction_of_gpu_memory_to_use=0.01 \
python3 ${BIN_DIR}/synthesize_e2e.py \
--config=vits_aishell3_ckpt_1.1.0/default.yaml \
--ckpt=vits_aishell3_ckpt_1.1.0/snapshot_iter_333000.pdz \
--phones_dict=vits_aishell3_ckpt_1.1.0/phone_id_map.txt \
--speaker_dict=vits_aishell3_ckpt_1.1.0/speaker_id_map.txt \
--output_dir=exp/default/test_e2e \
--text=${BIN_DIR}/../sentences.txt \
--add-blank=${add_blank}
```
-->
examples/aishell3/vits/conf/default.yaml 0 → 100644
浏览文件 @ 598eb1a5
# This configuration tested on 4 GPUs (V100) with 32GB GPU
# memory. It takes around 2 weeks to finish the training
# but 100k iters model should generate reasonable results.
###########################################################
# FEATURE EXTRACTION SETTING #
###########################################################
fs: 22050 # sr
n_fft: 1024 # FFT size (samples).
n_shift: 256 # Hop size (samples). 12.5ms
win_length: null # Window length (samples). 50ms
# If set to null, it will be the same as fft_size.
window: "hann" # Window function.
##########################################################
# TTS MODEL SETTING #
##########################################################
model:
# generator related
generator_type: vits_generator
generator_params:
hidden_channels: 192
global_channels: 256
segment_size: 32
text_encoder_attention_heads: 2
text_encoder_ffn_expand: 4
text_encoder_blocks: 6
text_encoder_positionwise_layer_type: "conv1d"
text_encoder_positionwise_conv_kernel_size: 3
text_encoder_positional_encoding_layer_type: "rel_pos"
text_encoder_self_attention_layer_type: "rel_selfattn"
text_encoder_activation_type: "swish"
text_encoder_normalize_before: True
text_encoder_dropout_rate: 0.1
text_encoder_positional_dropout_rate: 0.0
text_encoder_attention_dropout_rate: 0.1
use_macaron_style_in_text_encoder: True
use_conformer_conv_in_text_encoder: False
text_encoder_conformer_kernel_size: -1
decoder_kernel_size: 7
decoder_channels: 512
decoder_upsample_scales: [8, 8, 2, 2]
decoder_upsample_kernel_sizes: [16, 16, 4, 4]
decoder_resblock_kernel_sizes: [3, 7, 11]
decoder_resblock_dilations: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
use_weight_norm_in_decoder: True
posterior_encoder_kernel_size: 5
posterior_encoder_layers: 16
posterior_encoder_stacks: 1
posterior_encoder_base_dilation: 1
posterior_encoder_dropout_rate: 0.0
use_weight_norm_in_posterior_encoder: True
flow_flows: 4
flow_kernel_size: 5
flow_base_dilation: 1
flow_layers: 4
flow_dropout_rate: 0.0
use_weight_norm_in_flow: True
use_only_mean_in_flow: True
stochastic_duration_predictor_kernel_size: 3
stochastic_duration_predictor_dropout_rate: 0.5
stochastic_duration_predictor_flows: 4
stochastic_duration_predictor_dds_conv_layers: 3
# discriminator related
discriminator_type: hifigan_multi_scale_multi_period_discriminator
discriminator_params:
scales: 1
scale_downsample_pooling: "AvgPool1D"
scale_downsample_pooling_params:
kernel_size: 4
stride: 2
padding: 2
scale_discriminator_params:
in_channels: 1
out_channels: 1
kernel_sizes: [15, 41, 5, 3]
channels: 128
max_downsample_channels: 1024
max_groups: 16
bias: True
downsample_scales: [2, 2, 4, 4, 1]
nonlinear_activation: "leakyrelu"
nonlinear_activation_params:
negative_slope: 0.1
use_weight_norm: True
use_spectral_norm: False
follow_official_norm: False
periods: [2, 3, 5, 7, 11]
period_discriminator_params:
in_channels: 1
out_channels: 1
kernel_sizes: [5, 3]
channels: 32
downsample_scales: [3, 3, 3, 3, 1]
max_downsample_channels: 1024
bias: True
nonlinear_activation: "leakyrelu"
nonlinear_activation_params:
negative_slope: 0.1
use_weight_norm: True
use_spectral_norm: False
# others
sampling_rate: 22050 # needed in the inference for saving wav
cache_generator_outputs: True # whether to cache generator outputs in the training
###########################################################
# LOSS SETTING #
###########################################################
# loss function related
generator_adv_loss_params:
average_by_discriminators: False # whether to average loss value by #discriminators
loss_type: mse # loss type, "mse" or "hinge"
discriminator_adv_loss_params:
average_by_discriminators: False # whether to average loss value by #discriminators
loss_type: mse # loss type, "mse" or "hinge"
feat_match_loss_params:
average_by_discriminators: False # whether to average loss value by #discriminators
average_by_layers: False # whether to average loss value by #layers of each discriminator
include_final_outputs: True # whether to include final outputs for loss calculation
mel_loss_params:
fs: 22050 # must be the same as the training data
fft_size: 1024 # fft points
hop_size: 256 # hop size
win_length: null # window length
window: hann # window type
num_mels: 80 # number of Mel basis
fmin: 0 # minimum frequency for Mel basis
fmax: null # maximum frequency for Mel basis
log_base: null # null represent natural log
###########################################################
# ADVERSARIAL LOSS SETTING #
###########################################################
lambda_adv: 1.0 # loss scaling coefficient for adversarial loss
lambda_mel: 45.0 # loss scaling coefficient for Mel loss
lambda_feat_match: 2.0 # loss scaling coefficient for feat match loss
lambda_dur: 1.0 # loss scaling coefficient for duration loss
lambda_kl: 1.0 # loss scaling coefficient for KL divergence loss
# others
sampling_rate: 22050 # needed in the inference for saving wav
cache_generator_outputs: True # whether to cache generator outputs in the training
###########################################################
# DATA LOADER SETTING #
###########################################################
batch_size: 50 # Batch size.
num_workers: 4 # Number of workers in DataLoader.
##########################################################
# OPTIMIZER & SCHEDULER SETTING #
##########################################################
# optimizer setting for generator
generator_optimizer_params:
beta1: 0.8
beta2: 0.99
epsilon: 1.0e-9
weight_decay: 0.0
generator_scheduler: exponential_decay
generator_scheduler_params:
learning_rate: 2.0e-4
gamma: 0.999875
# optimizer setting for discriminator
discriminator_optimizer_params:
beta1: 0.8
beta2: 0.99
epsilon: 1.0e-9
weight_decay: 0.0
discriminator_scheduler: exponential_decay
discriminator_scheduler_params:
learning_rate: 2.0e-4
gamma: 0.999875
generator_first: False # whether to start updating generator first
##########################################################
# OTHER TRAINING SETTING #
##########################################################
num_snapshots: 10 # max number of snapshots to keep while training
train_max_steps: 350000 # Number of training steps. == total_iters / ngpus, total_iters = 1000000
save_interval_steps: 1000 # Interval steps to save checkpoint.
eval_interval_steps: 250 # Interval steps to evaluate the network.
seed: 777 # random seed number
examples/aishell3/vits/local/preprocess.sh 0 → 100755
浏览文件 @ 598eb1a5
#!/bin/bash
stage=0
stop_stage=100
config_path=$1
add_blank=$2
# copy from tts3/preprocess
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
# get durations from MFA's result
echo "Generate durations.txt from MFA results ..."
python3 ${MAIN_ROOT}/utils/gen_duration_from_textgrid.py \
--inputdir=./aishell3_alignment_tone \
--output durations.txt \
--config=${config_path}
fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# extract features
echo "Extract features ..."
python3 ${BIN_DIR}/preprocess.py \
--dataset=aishell3 \
--rootdir=~/datasets/data_aishell3/ \
--dumpdir=dump \
--dur-file=durations.txt \
--config=${config_path} \
--num-cpu=20 \
--cut-sil=True
fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
# get features' stats(mean and std)
echo "Get features' stats ..."
python3 ${MAIN_ROOT}/utils/compute_statistics.py \
--metadata=dump/train/raw/metadata.jsonl \
--field-name="feats"
fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# normalize and covert phone/speaker to id, dev and test should use train's stats
echo "Normalize ..."
python3 ${BIN_DIR}/normalize.py \
--metadata=dump/train/raw/metadata.jsonl \
--dumpdir=dump/train/norm \
--feats-stats=dump/train/feats_stats.npy \
--phones-dict=dump/phone_id_map.txt \
--speaker-dict=dump/speaker_id_map.txt \
--add-blank=${add_blank} \
--skip-wav-copy
python3 ${BIN_DIR}/normalize.py \
--metadata=dump/dev/raw/metadata.jsonl \
--dumpdir=dump/dev/norm \
--feats-stats=dump/train/feats_stats.npy \
--phones-dict=dump/phone_id_map.txt \
--speaker-dict=dump/speaker_id_map.txt \
--add-blank=${add_blank} \
--skip-wav-copy
python3 ${BIN_DIR}/normalize.py \
--metadata=dump/test/raw/metadata.jsonl \
--dumpdir=dump/test/norm \
--feats-stats=dump/train/feats_stats.npy \
--phones-dict=dump/phone_id_map.txt \
--speaker-dict=dump/speaker_id_map.txt \
--add-blank=${add_blank} \
--skip-wav-copy
fi
examples/aishell3/vits/local/synthesize.sh 0 → 100755
浏览文件 @ 598eb1a5
#!/bin/bash
config_path=$1
train_output_path=$2
ckpt_name=$3
stage=0
stop_stage=0
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
FLAGS_allocator_strategy=naive_best_fit \
FLAGS_fraction_of_gpu_memory_to_use=0.01 \
python3 ${BIN_DIR}/synthesize.py \
--config=${config_path} \
--ckpt=${train_output_path}/checkpoints/${ckpt_name} \
--phones_dict=dump/phone_id_map.txt \
--speaker_dict=dump/speaker_id_map.txt \
--test_metadata=dump/test/norm/metadata.jsonl \
--output_dir=${train_output_path}/test
fi
examples/aishell3/vits/local/synthesize_e2e.sh 0 → 100755
浏览文件 @ 598eb1a5
#!/bin/bash
config_path=$1
train_output_path=$2
ckpt_name=$3
add_blank=$4
stage=0
stop_stage=0
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
FLAGS_allocator_strategy=naive_best_fit \
FLAGS_fraction_of_gpu_memory_to_use=0.01 \
python3 ${BIN_DIR}/synthesize_e2e.py \
--config=${config_path} \
--ckpt=${train_output_path}/checkpoints/${ckpt_name} \
--phones_dict=dump/phone_id_map.txt \
--speaker_dict=dump/speaker_id_map.txt \
--spk_id=0 \
--output_dir=${train_output_path}/test_e2e \
--text=${BIN_DIR}/../sentences.txt \
--add-blank=${add_blank}
fi
examples/aishell3/vits/local/train.sh 0 → 100755
浏览文件 @ 598eb1a5
#!/bin/bash
config_path=$1
train_output_path=$2
# install monotonic_align
cd ${MAIN_ROOT}/paddlespeech/t2s/models/vits/monotonic_align
python3 setup.py build_ext --inplace
cd -
python3 ${BIN_DIR}/train.py \
--train-metadata=dump/train/norm/metadata.jsonl \
--dev-metadata=dump/dev/norm/metadata.jsonl \
--config=${config_path} \
--output-dir=${train_output_path} \
--ngpu=4 \
--phones-dict=dump/phone_id_map.txt \
--speaker-dict=dump/speaker_id_map.txt
examples/aishell3/vits/path.sh 0 → 100755
浏览文件 @ 598eb1a5
#!/bin/bash
export MAIN_ROOT=`realpath ${PWD}/../../../`
export PATH=${MAIN_ROOT}:${MAIN_ROOT}/utils:${PATH}
export LC_ALL=C
export PYTHONDONTWRITEBYTECODE=1
# Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
export PYTHONIOENCODING=UTF-8
export PYTHONPATH=${MAIN_ROOT}:${PYTHONPATH}
MODEL=vits
export BIN_DIR=${MAIN_ROOT}/paddlespeech/t2s/exps/${MODEL}
\ No newline at end of file
examples/aishell3/vits/run.sh 0 → 100755
浏览文件 @ 598eb1a5
#!/bin/bash
set -e
source path.sh
gpus=0,1,2,3
stage=0
stop_stage=100
conf_path=conf/default.yaml
train_output_path=exp/default
ckpt_name=snapshot_iter_153.pdz
add_blank=true
# with the following command, you can choose the stage range you want to run
# such as `./run.sh --stage 0 --stop-stage 0`
# this can not be mixed use with `$1`, `$2` ...
source ${MAIN_ROOT}/utils/parse_options.sh || exit 1
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
# prepare data
./local/preprocess.sh ${conf_path} ${add_blank}|| exit -1
fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# train model, all `ckpt` under `train_output_path/checkpoints/` dir
CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path} || exit -1
fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1
fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize_e2e.sh ${conf_path} ${train_output_path} ${ckpt_name} ${add_blank}|| exit -1
fi
examples/aishell3_vctk/ernie_sat/run.sh
浏览文件 @ 598eb1a5
......@@ -3,7 +3,7 @@
set -e
source path.sh
gpus=0,1
gpus=0,1,2,3,4,5,6,7
stage=0
stop_stage=100
......
examples/csmsc/tts2/run.sh
浏览文件 @ 598eb1a5
......@@ -46,8 +46,8 @@ fi
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
# install paddle2onnx
version=$(echo `pip list |grep "paddle2onnx"` |awk -F" " '{print $2}')
if [[ -z "$version" || ${version} != '0.9.8' ]]; then
pip install paddle2onnx==0.9.8
if [[ -z "$version" || ${version} != '1.0.0' ]]; then
pip install paddle2onnx==1.0.0
fi
./local/paddle2onnx.sh ${train_output_path} inference inference_onnx speedyspeech_csmsc
# considering the balance between speed and quality, we recommend that you use hifigan as vocoder
......
examples/csmsc/tts3/run.sh
浏览文件 @ 598eb1a5
......@@ -46,8 +46,8 @@ fi
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
# install paddle2onnx
version=$(echo `pip list |grep "paddle2onnx"` |awk -F" " '{print $2}')
if [[ -z "$version" || ${version} != '0.9.8' ]]; then
pip install paddle2onnx==0.9.8
if [[ -z "$version" || ${version} != '1.0.0' ]]; then
pip install paddle2onnx==1.0.0
fi
./local/paddle2onnx.sh ${train_output_path} inference inference_onnx fastspeech2_csmsc
# considering the balance between speed and quality, we recommend that you use hifigan as vocoder
......
examples/csmsc/tts3/run_cnndecoder.sh
浏览文件 @ 598eb1a5
......@@ -59,8 +59,8 @@ fi
if [ ${stage} -le 7 ] && [ ${stop_stage} -ge 7 ]; then
# install paddle2onnx
version=$(echo `pip list |grep "paddle2onnx"` |awk -F" " '{print $2}')
if [[ -z "$version" || ${version} != '0.9.8' ]]; then
pip install paddle2onnx==0.9.8
if [[ -z "$version" || ${version} != '1.0.0' ]]; then
pip install paddle2onnx==1.0.0
fi
./local/paddle2onnx.sh ${train_output_path} inference inference_onnx fastspeech2_csmsc
# considering the balance between speed and quality, we recommend that you use hifigan as vocoder
......@@ -79,8 +79,8 @@ fi
if [ ${stage} -le 9 ] && [ ${stop_stage} -ge 9 ]; then
# install paddle2onnx
version=$(echo `pip list |grep "paddle2onnx"` |awk -F" " '{print $2}')
if [[ -z "$version" || ${version} != '0.9.8' ]]; then
pip install paddle2onnx==0.9.8
if [[ -z "$version" || ${version} != '1.0.0' ]]; then
pip install paddle2onnx==1.0.0
fi
# streaming acoustic model
./local/paddle2onnx.sh ${train_output_path} inference_streaming inference_onnx_streaming fastspeech2_csmsc_am_encoder_infer
......
examples/csmsc/vits/run.sh
浏览文件 @ 598eb1a5
......@@ -3,7 +3,7 @@
set -e
source path.sh
gpus=0,1
gpus=0,1,2,3
stage=0
stop_stage=100
......
examples/hey_snips/kws0/conf/mdtc.yaml
浏览文件 @ 598eb1a5
......@@ -3,7 +3,7 @@
# Data #
###########################################
dataset: 'paddlespeech.audio.datasets:HeySnips'
data_dir: '/PATH/TO/DATA/hey_snips_research_6k_en_train_eval_clean_ter'
data_dir: '../tests/hey_snips_research_6k_en_train_eval_clean_ter'
############################################
# Network Architecture #
......@@ -46,4 +46,4 @@ num_workers: 16
checkpoint: './checkpoint/epoch_100/model.pdparams'
score_file: './scores.txt'
stats_file: './stats.0.txt'
img_file: './det.png'
\ No newline at end of file
img_file: './det.png'
examples/iwslt2012/punc0/README.md
浏览文件 @ 598eb1a5
......@@ -18,12 +18,62 @@
./run.sh --stage 3 --stop-stage 3
```
## Pretrained Model
The pretrained model can be downloaded here [ernie_linear_p3_iwslt2012_zh_ckpt_0.1.1.zip](https://paddlespeech.bj.bcebos.com/text/ernie_linear_p3_iwslt2012_zh_ckpt_0.1.1.zip).
The pretrained model can be downloaded here:
[ernie_linear_p3_iwslt2012_zh_ckpt_0.1.1.zip](https://paddlespeech.bj.bcebos.com/text/ernie_linear_p3_iwslt2012_zh_ckpt_0.1.1.zip)
[ernie-3.0-base.tar.gz](https://paddlespeech.bj.bcebos.com/punc_restore/ernie-3.0-base.tar.gz)
[ernie-3.0-medium.tar.gz](https://paddlespeech.bj.bcebos.com/punc_restore/ernie-3.0-medium.tar.gz)
[ernie-3.0-micro.tar.gz](https://paddlespeech.bj.bcebos.com/punc_restore/ernie-3.0-micro.tar.gz)
[ernie-mini.tar.gz](https://paddlespeech.bj.bcebos.com/punc_restore/ernie-mini.tar.gz)
[ernie-nano.tar.gz](https://paddlespeech.bj.bcebos.com/punc_restore/ernie-nano.tar.gz)
[ernie-tiny.tar.gz](https://paddlespeech.bj.bcebos.com/punc_restore/ernie-tiny.tar.gz)
### Test Result
- Ernie
- Ernie 1.0
| |COMMA | PERIOD | QUESTION | OVERALL|
|:-----:|:-----:|:-----:|:-----:|:-----:|
|Precision |0.510955 |0.526462 |0.820755 |0.619391|
|Recall |0.517433 |0.564179 |0.861386 |0.647666|
|F1 |0.514173 |0.544669 |0.840580 |0.633141|
- Ernie-tiny
| |COMMA | PERIOD | QUESTION | OVERALL|
|:-----:|:-----:|:-----:|:-----:|:-----:|
|Precision |0.733177 |0.721448 |0.754717 |0.736447|
|Recall |0.380740 |0.524646 |0.733945 |0.546443|
|F1 |0.501204 |0.607506 |0.744186 |0.617632|
- Ernie-3.0-base-zh
| |COMMA | PERIOD | QUESTION | OVERALL|
|:-----:|:-----:|:-----:|:-----:|:-----:|
|Precision |0.805947 |0.764160 |0.858491 |0.809532|
|Recall |0.399070 |0.567978 |0.850467 |0.605838|
|F1 |0.533817 |0.651623 |0.854460 |0.679967|
- Ernie-3.0-medium-zh
| |COMMA | PERIOD | QUESTION | OVERALL|
|:-----:|:-----:|:-----:|:-----:|:-----:|
|Precision |0.730829 |0.699164 |0.707547 |0.712514|
|Recall |0.388196 |0.533286 |0.797872 |0.573118|
|F1 |0.507058 |0.605062 |0.750000 |0.620707|
- Ernie-3.0-mini-zh
| |COMMA | PERIOD | QUESTION | OVERALL|
|:-----:|:-----:|:-----:|:-----:|:-----:|
|Precision |0.757433 |0.708449 |0.707547 |0.724477|
|Recall |0.355752 |0.506977 |0.735294 |0.532674|
|F1 |0.484121 |0.591015 |0.721154 |0.598763|
- Ernie-3.0-micro-zh
| |COMMA | PERIOD | QUESTION | OVERALL|
|:-----:|:-----:|:-----:|:-----:|:-----:|
|Precision |0.733959 |0.679666 |0.726415 |0.713347|
|Recall |0.332742 |0.483487 |0.712963 |0.509731|
|F1 |0.457896 |0.565033 |0.719626 |0.580852|
- Ernie-3.0-nano-zh
| |COMMA | PERIOD | QUESTION | OVERALL|
|:-----:|:-----:|:-----:|:-----:|:-----:|
|Precision |0.693271 |0.682451 |0.754717 |0.710146|
|Recall |0.327784 |0.491968 |0.666667 |0.495473|
|F1 |0.445114 |0.571762 |0.707965 |0.574947|
examples/iwslt2012/punc0/local/preprocess.py
浏览文件 @ 598eb1a5
import argparse
import os
def process_sentence(line):
if line == '': return ''
res = line[0]
for i in range(1, len(line)):
res += (' ' + line[i])
return res
if line == '':
return ''
res = line[0]
for i in range(1, len(line)):
res += (' ' + line[i])
return res
if __name__ == "__main__":
paser = argparse.ArgumentParser(description = "Input filename")
paser.add_argument('-input_file')
paser.add_argument('-output_file')
sentence_cnt = 0
args = paser.parse_args()
with open(args.input_file, 'r') as f:
with open(args.output_file, 'w') as write_f:
while True:
line = f.readline()
if line:
sentence_cnt += 1
write_f.write(process_sentence(line))
else:
break
print('preprocess over')
print('total sentences number:', sentence_cnt)
paser = argparse.ArgumentParser(description="Input filename")
paser.add_argument('-input_file')
paser.add_argument('-output_file')
sentence_cnt = 0
args = paser.parse_args()
with open(args.input_file, 'r') as f:
with open(args.output_file, 'w') as write_f:
while True:
line = f.readline()
if line:
sentence_cnt += 1
write_f.write(process_sentence(line))
else:
break
print('preprocess over')
print('total sentences number:', sentence_cnt)
examples/ljspeech/tts3/run.sh
浏览文件 @ 598eb1a5
......@@ -46,8 +46,8 @@ fi
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
# install paddle2onnx
version=$(echo `pip list |grep "paddle2onnx"` |awk -F" " '{print $2}')
if [[ -z "$version" || ${version} != '0.9.8' ]]; then
pip install paddle2onnx==0.9.8
if [[ -z "$version" || ${version} != '1.0.0' ]]; then
pip install paddle2onnx==1.0.0
fi
./local/paddle2onnx.sh ${train_output_path} inference inference_onnx fastspeech2_ljspeech
# considering the balance between speed and quality, we recommend that you use hifigan as vocoder
......
examples/other/g2p/README.md
浏览文件 @ 598eb1a5
......@@ -9,16 +9,19 @@ We use `WER` as an evaluation criterion.
Run the command below to get the results of the test.
```bash
cd ../../../tools
bash extras/install_sclite.sh
cd -
./run.sh
```
The `avg WER` of g2p is: 0.024169315564825305
The `avg WER` of g2p is: 0.024075726733983775
```text
,--------------------------------------------------------------------.
| ./exp/g2p/text.g2p |
|--------------------------------------------------------------------|
| SPKR | # Snt # Wrd | Corr Sub Del Ins Err S.Err |
| Sum/Avg| 9996 299181 | 97.6 2.4 0.0 0.0 2.4 49.2 |
| Sum/Avg| 9996 299181 | 97.6 2.4 0.0 0.0 2.4 49.0 |
`--------------------------------------------------------------------'
```
examples/other/g2p/compare_badcase.py 0 → 100644
浏览文件 @ 598eb1a5
# -*- encoding:utf-8 -*-
import re
import sys
'''
@arthur: david_95
Assum you executed g2p test twice, the WER rate have some gap, you would like to see what sentences error cause your rate up.
so you may get test result ( exp/g2p )into two directories, as exp/prefolder and exp/curfolder
run this program as "python compare_badcase.py prefolder curfolder"
then you will get diffrences between two run, uuid, phonetics, chinese samples
examples: python compare_badcase.py exp/g2p_laotouzi exp/g2p
in this example: exp/g2p_laotouzi and exp/g2p are two folders with two g2p tests result
'''
def compare(prefolder, curfolder):
'''
compare file of text.g2p.pra in two folders
result P1 will be prefolder ; P2 will be curfolder, just about the sequence you input in argvs
'''
linecnt = 0
pre_block = []
cur_block = []
zh_lines = []
with open(prefolder + "/text.g2p.pra", "r") as pre_file, open(
curfolder + "/text.g2p.pra", "r") as cur_file:
for pre_line, cur_line in zip(pre_file, cur_file):
linecnt += 1
if linecnt < 11: #skip non-data head in files
continue
else:
pre_block.append(pre_line.strip())
cur_block.append(cur_line.strip())
if pre_line.strip().startswith(
"Eval:") and pre_line.strip() != cur_line.strip():
uuid = pre_block[-5].replace("id: (baker_", "").replace(")",
"")
with open("data/g2p/text", 'r') as txt:
conlines = txt.readlines()
for line in conlines:
if line.strip().startswith(uuid.strip()):
print(line)
zh_lines.append(re.sub(r"#[1234]", "", line))
break
print("*" + cur_block[-3]) # ref
print("P1 " + pre_block[-2])
print("P2 " + cur_block[-2])
print("P1 " + pre_block[-1])
print("P2 " + cur_block[-1] + "\n\n")
pre_block = []
cur_block = []
print("\n")
print(str.join("\n", zh_lines))
if __name__ == '__main__':
assert len(
sys.argv) == 3, "Usage: python compare_badcase.py %prefolder %curfolder"
compare(sys.argv[1], sys.argv[2])
examples/other/tn/README.md
浏览文件 @ 598eb1a5
......@@ -5,6 +5,9 @@ We use `CER` as an evaluation criterion.
## Start
Run the command below to get the results of the test.
```bash
cd ../../../tools
bash extras/install_sclite.sh
cd -
./run.sh
```
The `avg CER` of text normalization is: 0.00730093543235227
......
examples/other/tts_finetune/tts3/finetune.py
浏览文件 @ 598eb1a5
......@@ -17,15 +17,14 @@ from pathlib import Path
from typing import Union
import yaml
from paddle import distributed as dist
from yacs.config import CfgNode
from paddlespeech.t2s.exps.fastspeech2.train import train_sp
from local.check_oov import get_check_result
from local.extract import extract_feature
from local.label_process import get_single_label
from local.prepare_env import generate_finetune_env
from paddle import distributed as dist
from yacs.config import CfgNode
from paddlespeech.t2s.exps.fastspeech2.train import train_sp
from utils.gen_duration_from_textgrid import gen_duration_from_textgrid
DICT_EN = 'tools/aligner/cmudict-0.7b'
......
examples/vctk/ernie_sat/run.sh
浏览文件 @ 598eb1a5
......@@ -3,7 +3,7 @@
set -e
source path.sh
gpus=0,1
gpus=0,1,2,3,4,5,6,7
stage=0
stop_stage=100
......
examples/vctk/tts3/run.sh
浏览文件 @ 598eb1a5
......@@ -44,8 +44,8 @@ fi
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
# install paddle2onnx
version=$(echo `pip list |grep "paddle2onnx"` |awk -F" " '{print $2}')
if [[ -z "$version" || ${version} != '0.9.8' ]]; then
pip install paddle2onnx==0.9.8
if [[ -z "$version" || ${version} != '1.0.0' ]]; then
pip install paddle2onnx==1.0.0
fi
./local/paddle2onnx.sh ${train_output_path} inference inference_onnx fastspeech2_vctk
# considering the balance between speed and quality, we recommend that you use hifigan as vocoder
......
examples/voxceleb/sv0/RESULT.md
浏览文件 @ 598eb1a5
......@@ -5,3 +5,7 @@
| Model | Number of Params | Release | Config | dim | Test set | Cosine | Cosine + S-Norm |
| --- | --- | --- | --- | --- | --- | --- | ---- |
| ECAPA-TDNN | 85M | 0.2.1 | conf/ecapa_tdnn.yaml | 192 | test | 0.8188 | 0.7815|
> [SpeechBrain result](https://github.com/speechbrain/speechbrain/tree/develop/recipes/VoxCeleb/SpeakerRec#speaker-verification-using-ecapa-tdnn-embeddings):
> EER = 0.90% (voxceleb1 + voxceleb2) without s-norm
> EER = 0.80% (voxceleb1 + voxceleb2) with s-norm.
examples/zh_en_tts/tts3/run.sh
浏览文件 @ 598eb1a5
......@@ -47,8 +47,8 @@ fi
if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
# install paddle2onnx
version=$(echo `pip list |grep "paddle2onnx"` |awk -F" " '{print $2}')
if [[ -z "$version" || ${version} != '0.9.8' ]]; then
pip install paddle2onnx==0.9.8
if [[ -z "$version" || ${version} != '1.0.0' ]]; then
pip install paddle2onnx==1.0.0
fi
./local/paddle2onnx.sh ${train_output_path} inference inference_onnx fastspeech2_mix
# considering the balance between speed and quality, we recommend that you use hifigan as vocoder
......
paddlespeech/__init__.py
浏览文件 @ 598eb1a5
......@@ -14,5 +14,3 @@
import _locale
_locale._getdefaultlocale = (lambda *args: ['en_US', 'utf8'])
paddlespeech/audio/__init__.py
浏览文件 @ 598eb1a5
......@@ -14,12 +14,12 @@
from . import compliance
from . import datasets
from . import features
from . import text
from . import transform
from . import streamdata
from . import functional
from . import io
from . import metric
from . import sox_effects
from . import streamdata
from . import text
from . import transform
from .backends import load
from .backends import save
paddlespeech/audio/streamdata/__init__.py
浏览文件 @ 598eb1a5
......@@ -4,67 +4,66 @@
# Modified from https://github.com/webdataset/webdataset
#
# flake8: noqa
from .cache import (
cached_tarfile_samples,
cached_tarfile_to_samples,
lru_cleanup,
pipe_cleaner,
)
from .compat import WebDataset, WebLoader, FluidWrapper
from .extradatasets import MockDataset, with_epoch, with_length
from .filters import (
associate,
batched,
decode,
detshuffle,
extract_keys,
getfirst,
info,
map,
map_dict,
map_tuple,
pipelinefilter,
rename,
rename_keys,
audio_resample,
select,
shuffle,
slice,
to_tuple,
transform_with,
unbatched,
xdecode,
audio_data_filter,
audio_tokenize,
audio_resample,
audio_compute_fbank,
audio_spec_aug,
sort,
audio_padding,
audio_cmvn,
placeholder,
)
from .handlers import (
ignore_and_continue,
ignore_and_stop,
reraise_exception,
warn_and_continue,
warn_and_stop,
)
from .cache import cached_tarfile_samples
from .cache import cached_tarfile_to_samples
from .cache import lru_cleanup
from .cache import pipe_cleaner
from .compat import FluidWrapper
from .compat import WebDataset
from .compat import WebLoader
from .extradatasets import MockDataset
from .extradatasets import with_epoch
from .extradatasets import with_length
from .filters import associate
from .filters import audio_cmvn
from .filters import audio_compute_fbank
from .filters import audio_data_filter
from .filters import audio_padding
from .filters import audio_resample
from .filters import audio_spec_aug
from .filters import audio_tokenize
from .filters import batched
from .filters import decode
from .filters import detshuffle
from .filters import extract_keys
from .filters import getfirst
from .filters import info
from .filters import map
from .filters import map_dict
from .filters import map_tuple
from .filters import pipelinefilter
from .filters import placeholder
from .filters import rename
from .filters import rename_keys
from .filters import select
from .filters import shuffle
from .filters import slice
from .filters import sort
from .filters import to_tuple
from .filters import transform_with
from .filters import unbatched
from .filters import xdecode
from .handlers import ignore_and_continue
from .handlers import ignore_and_stop
from .handlers import reraise_exception
from .handlers import warn_and_continue
from .handlers import warn_and_stop
from .mix import RandomMix
from .mix import RoundRobin
from .pipeline import DataPipeline
from .shardlists import (
MultiShardSample,
ResampledShards,
SimpleShardList,
non_empty,
resampled,
shardspec,
single_node_only,
split_by_node,
split_by_worker,
)
from .tariterators import tarfile_samples, tarfile_to_samples
from .utils import PipelineStage, repeatedly
from .writer import ShardWriter, TarWriter, numpy_dumps
from .mix import RandomMix, RoundRobin
from .shardlists import MultiShardSample
from .shardlists import non_empty
from .shardlists import resampled
from .shardlists import ResampledShards
from .shardlists import shardspec
from .shardlists import SimpleShardList
from .shardlists import single_node_only
from .shardlists import split_by_node
from .shardlists import split_by_worker
from .tariterators import tarfile_samples
from .tariterators import tarfile_to_samples
from .utils import PipelineStage
from .utils import repeatedly
from .writer import numpy_dumps
from .writer import ShardWriter
from .writer import TarWriter
paddlespeech/audio/streamdata/autodecode.py
浏览文件 @ 598eb1a5
......@@ -5,18 +5,19 @@
# See the LICENSE file for licensing terms (BSD-style).
# Modified from https://github.com/webdataset/webdataset
#
"""Automatically decode webdataset samples."""
import io, json, os, pickle, re, tempfile
import io
import json
import os
import pickle
import re
import tempfile
from functools import partial
import numpy as np
"""Extensions passed on to the image decoder."""
image_extensions = "jpg jpeg png ppm pgm pbm pnm".split()
################################################################
# handle basic datatypes
################################################################
......@@ -128,7 +129,7 @@ def call_extension_handler(key, data, f, extensions):
target = target.split(".")
if len(target) > len(extension):
continue
if extension[-len(target) :] == target:
if extension[-len(target):] == target:
return f(data)
return None
......@@ -268,7 +269,6 @@ def imagehandler(imagespec, extensions=image_extensions):
################################################################
# torch video
################################################################
'''
def torch_video(key, data):
"""Decode video using the torchvideo library.
......@@ -289,7 +289,6 @@ def torch_video(key, data):
return torchvision.io.read_video(fname, pts_unit="sec")
'''
################################################################
# paddlespeech.audio
################################################################
......@@ -359,7 +358,6 @@ def gzfilter(key, data):
# decode entire training amples
################################################################
default_pre_handlers = [gzfilter]
default_post_handlers = [basichandlers]
......@@ -387,7 +385,8 @@ class Decoder:
pre = default_pre_handlers
if post is None:
post = default_post_handlers
assert all(callable(h) for h in handlers), f"one of {handlers} not callable"
assert all(callable(h)
for h in handlers), f"one of {handlers} not callable"
assert all(callable(h) for h in pre), f"one of {pre} not callable"
assert all(callable(h) for h in post), f"one of {post} not callable"
self.handlers = pre + handlers + post
......
paddlespeech/audio/streamdata/cache.py
浏览文件 @ 598eb1a5
......@@ -2,7 +2,10 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
# See the LICENSE file for licensing terms (BSD-style).
# Modified from https://github.com/webdataset/webdataset
import itertools, os, random, re, sys
import os
import random
import re
import sys
from urllib.parse import urlparse
from . import filters
......@@ -40,7 +43,7 @@ def lru_cleanup(cache_dir, cache_size, keyfn=os.path.getctime, verbose=False):
os.remove(fname)
def download(url, dest, chunk_size=1024 ** 2, verbose=False):
def download(url, dest, chunk_size=1024**2, verbose=False):
"""Download a file from `url` to `dest`."""
temp = dest + f".temp{os.getpid()}"
with gopen.gopen(url) as stream:
......@@ -65,12 +68,11 @@ def pipe_cleaner(spec):
def get_file_cached(
spec,
cache_size=-1,
cache_dir=None,
url_to_name=pipe_cleaner,
verbose=False,
):
spec,
cache_size=-1,
cache_dir=None,
url_to_name=pipe_cleaner,
verbose=False, ):
if cache_size == -1:
cache_size = default_cache_size
if cache_dir is None:
......@@ -107,15 +109,14 @@ verbose_cache = int(os.environ.get("WDS_VERBOSE_CACHE", "0"))
def cached_url_opener(
data,
handler=reraise_exception,
cache_size=-1,
cache_dir=None,
url_to_name=pipe_cleaner,
validator=check_tar_format,
verbose=False,
always=False,
):
data,
handler=reraise_exception,
cache_size=-1,
cache_dir=None,
url_to_name=pipe_cleaner,
validator=check_tar_format,
verbose=False,
always=False, ):
"""Given a stream of url names (packaged in `dict(url=url)`), yield opened streams."""
verbose = verbose or verbose_cache
for sample in data:
......@@ -132,8 +133,7 @@ def cached_url_opener(
cache_size=cache_size,
cache_dir=cache_dir,
url_to_name=url_to_name,
verbose=verbose,
)
verbose=verbose, )
if verbose:
print("# opening %s" % dest, file=sys.stderr)
assert os.path.exists(dest)
......@@ -143,9 +143,8 @@ def cached_url_opener(
data = f.read(200)
os.remove(dest)
raise ValueError(
"%s (%s) is not a tar archive, but a %s, contains %s"
% (dest, url, ftype, repr(data))
)
"%s (%s) is not a tar archive, but a %s, contains %s" %
(dest, url, ftype, repr(data)))
try:
stream = open(dest, "rb")
sample.update(stream=stream)
......@@ -158,7 +157,7 @@ def cached_url_opener(
continue
raise exn
except Exception as exn:
exn.args = exn.args + (url,)
exn.args = exn.args + (url, )
if handler(exn):
continue
else:
......@@ -166,14 +165,13 @@ def cached_url_opener(
def cached_tarfile_samples(
src,
handler=reraise_exception,
cache_size=-1,
cache_dir=None,
verbose=False,
url_to_name=pipe_cleaner,
always=False,
):
src,
handler=reraise_exception,
cache_size=-1,
cache_dir=None,
verbose=False,
url_to_name=pipe_cleaner,
always=False, ):
streams = cached_url_opener(
src,
handler=handler,
......@@ -181,8 +179,7 @@ def cached_tarfile_samples(
cache_dir=cache_dir,
verbose=verbose,
url_to_name=url_to_name,
always=always,
)
always=always, )
samples = tar_file_and_group_expander(streams, handler=handler)
return samples
......
paddlespeech/audio/streamdata/compat.py
浏览文件 @ 598eb1a5
......@@ -2,17 +2,17 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
# See the LICENSE file for licensing terms (BSD-style).
# Modified from https://github.com/webdataset/webdataset
from dataclasses import dataclass
from itertools import islice
from typing import List
import braceexpand, yaml
import yaml
from . import autodecode
from . import cache, filters, shardlists, tariterators
from . import cache
from . import filters
from . import shardlists
from . import tariterators
from .filters import reraise_exception
from .paddle_utils import DataLoader
from .paddle_utils import IterableDataset
from .pipeline import DataPipeline
from .paddle_utils import DataLoader, IterableDataset
class FluidInterface:
......@@ -26,7 +26,8 @@ class FluidInterface:
return self.compose(filters.unbatched())
def listed(self, batchsize, partial=True):
return self.compose(filters.batched(), batchsize=batchsize, collation_fn=None)
return self.compose(
filters.batched(), batchsize=batchsize, collation_fn=None)
def unlisted(self):
return self.compose(filters.unlisted())
......@@ -43,9 +44,19 @@ class FluidInterface:
def map(self, f, handler=reraise_exception):
return self.compose(filters.map(f, handler=handler))
def decode(self, *args, pre=None, post=None, only=None, partial=False, handler=reraise_exception):
handlers = [autodecode.ImageHandler(x) if isinstance(x, str) else x for x in args]
decoder = autodecode.Decoder(handlers, pre=pre, post=post, only=only, partial=partial)
def decode(self,
*args,
pre=None,
post=None,
only=None,
partial=False,
handler=reraise_exception):
handlers = [
autodecode.ImageHandler(x) if isinstance(x, str) else x
for x in args
]
decoder = autodecode.Decoder(
handlers, pre=pre, post=post, only=only, partial=partial)
return self.map(decoder, handler=handler)
def map_dict(self, handler=reraise_exception, **kw):
......@@ -80,12 +91,12 @@ class FluidInterface:
def audio_data_filter(self, *args, **kw):
return self.compose(filters.audio_data_filter(*args, **kw))
def audio_tokenize(self, *args, **kw):
return self.compose(filters.audio_tokenize(*args, **kw))
def resample(self, *args, **kw):
return self.compose(filters.resample(*args, **kw))
return self.compose(filters.resample(*args, **kw))
def audio_compute_fbank(self, *args, **kw):
return self.compose(filters.audio_compute_fbank(*args, **kw))
......@@ -102,27 +113,28 @@ class FluidInterface:
def audio_cmvn(self, cmvn_file):
return self.compose(filters.audio_cmvn(cmvn_file))
class WebDataset(DataPipeline, FluidInterface):
"""Small fluid-interface wrapper for DataPipeline."""
def __init__(
self,
urls,
handler=reraise_exception,
resampled=False,
repeat=False,
shardshuffle=None,
cache_size=0,
cache_dir=None,
detshuffle=False,
nodesplitter=shardlists.single_node_only,
verbose=False,
):
self,
urls,
handler=reraise_exception,
resampled=False,
repeat=False,
shardshuffle=None,
cache_size=0,
cache_dir=None,
detshuffle=False,
nodesplitter=shardlists.single_node_only,
verbose=False, ):
super().__init__()
if isinstance(urls, IterableDataset):
assert not resampled
self.append(urls)
elif isinstance(urls, str) and (urls.endswith(".yaml") or urls.endswith(".yml")):
elif isinstance(urls, str) and (urls.endswith(".yaml") or
urls.endswith(".yml")):
with (open(urls)) as stream:
spec = yaml.safe_load(stream)
assert "datasets" in spec
......@@ -152,9 +164,7 @@ class WebDataset(DataPipeline, FluidInterface):
handler=handler,
verbose=verbose,
cache_size=cache_size,
cache_dir=cache_dir,
)
)
cache_dir=cache_dir, ))
class FluidWrapper(DataPipeline, FluidInterface):
......
paddlespeech/audio/streamdata/extradatasets.py
浏览文件 @ 598eb1a5
......@@ -5,20 +5,10 @@
# See the LICENSE file for licensing terms (BSD-style).
# Modified from https://github.com/webdataset/webdataset
#
"""Train PyTorch models directly from POSIX tar archive.
Code works locally or over HTTP connections.
"""
import itertools as itt
import os
import random
import sys
import braceexpand
from . import utils
from .paddle_utils import IterableDataset
from .utils import PipelineStage
......@@ -63,8 +53,7 @@ class repeatedly(IterableDataset, PipelineStage):
return utils.repeatedly(
source,
nepochs=self.nepochs,
nbatches=self.nbatches,
)
nbatches=self.nbatches, )
class with_epoch(IterableDataset):
......
paddlespeech/audio/streamdata/filters.py
浏览文件 @ 598eb1a5
......@@ -3,7 +3,6 @@
# This file is part of the WebDataset library.
# See the LICENSE file for licensing terms (BSD-style).
#
# Modified from https://github.com/webdataset/webdataset
# Modified from wenet(https://github.com/wenet-e2e/wenet)
"""A collection of iterators for data transformations.
......@@ -12,28 +11,29 @@ These functions are plain iterator functions. You can find curried versions
in webdataset.filters, and you can find IterableDataset wrappers in
webdataset.processing.
"""
import io
from fnmatch import fnmatch
import itertools
import os
import random
import re
import itertools, os, random, sys, time
from functools import reduce, wraps
import sys
import time
from fnmatch import fnmatch
from functools import reduce
import numpy as np
import paddle
from . import autodecode
from . import utils
from .paddle_utils import PaddleTensor
from .utils import PipelineStage
from . import utils
from .. import backends
from ..compliance import kaldi
import paddle
from ..transform.cmvn import GlobalCMVN
from ..utils.tensor_utils import pad_sequence
from ..transform.spec_augment import time_warp
from ..transform.spec_augment import time_mask
from ..transform.spec_augment import freq_mask
from ..transform.spec_augment import time_mask
from ..transform.spec_augment import time_warp
from ..utils.tensor_utils import pad_sequence
from .utils import PipelineStage
class FilterFunction(object):
"""Helper class for currying pipeline stages.
......@@ -159,10 +159,12 @@ def transform_with(sample, transformers):
result[i] = f(sample[i])
return result
###
# Iterators
###
def _info(data, fmt=None, n=3, every=-1, width=50, stream=sys.stderr, name=""):
"""Print information about the samples that are passing through.
......@@ -278,10 +280,16 @@ def _log_keys(data, logfile=None):
log_keys = pipelinefilter(_log_keys)
def _minedecode(x):
if isinstance(x, str):
return autodecode.imagehandler(x)
else:
return x
def _decode(data, *args, handler=reraise_exception, **kw):
"""Decode data based on the decoding functions given as arguments."""
decoder = lambda x: autodecode.imagehandler(x) if isinstance(x, str) else x
decoder = _minedecode
handlers = [decoder(x) for x in args]
f = autodecode.Decoder(handlers, **kw)
......@@ -325,15 +333,24 @@ def _rename(data, handler=reraise_exception, keep=True, **kw):
for sample in data:
try:
if not keep:
yield {k: getfirst(sample, v, missing_is_error=True) for k, v in kw.items()}
yield {
k: getfirst(sample, v, missing_is_error=True)
for k, v in kw.items()
}
else:
def listify(v):
return v.split(";") if isinstance(v, str) else v
to_be_replaced = {x for v in kw.values() for x in listify(v)}
result = {k: v for k, v in sample.items() if k not in to_be_replaced}
result.update({k: getfirst(sample, v, missing_is_error=True) for k, v in kw.items()})
result = {
k: v
for k, v in sample.items() if k not in to_be_replaced
}
result.update({
k: getfirst(sample, v, missing_is_error=True)
for k, v in kw.items()
})
yield result
except Exception as exn:
if handler(exn):
......@@ -381,7 +398,11 @@ def _map_dict(data, handler=reraise_exception, **kw):
map_dict = pipelinefilter(_map_dict)
def _to_tuple(data, *args, handler=reraise_exception, missing_is_error=True, none_is_error=None):
def _to_tuple(data,
*args,
handler=reraise_exception,
missing_is_error=True,
none_is_error=None):
"""Convert dict samples to tuples."""
if none_is_error is None:
none_is_error = missing_is_error
......@@ -390,7 +411,10 @@ def _to_tuple(data, *args, handler=reraise_exception, missing_is_error=True, non
for sample in data:
try:
result = tuple([getfirst(sample, f, missing_is_error=missing_is_error) for f in args])
result = tuple([
getfirst(sample, f, missing_is_error=missing_is_error)
for f in args
])
if none_is_error and any(x is None for x in result):
raise ValueError(f"to_tuple {args} got {sample.keys()}")
yield result
......@@ -463,19 +487,28 @@ rsample = pipelinefilter(_rsample)
slice = pipelinefilter(itertools.islice)
def _extract_keys(source, *patterns, duplicate_is_error=True, ignore_missing=False):
def _extract_keys(source,
*patterns,
duplicate_is_error=True,
ignore_missing=False):
for sample in source:
result = []
for pattern in patterns:
pattern = pattern.split(";") if isinstance(pattern, str) else pattern
matches = [x for x in sample.keys() if any(fnmatch("." + x, p) for p in pattern)]
pattern = pattern.split(";") if isinstance(pattern,
str) else pattern
matches = [
x for x in sample.keys()
if any(fnmatch("." + x, p) for p in pattern)
]
if len(matches) == 0:
if ignore_missing:
continue
else:
raise ValueError(f"Cannot find {pattern} in sample keys {sample.keys()}.")
raise ValueError(
f"Cannot find {pattern} in sample keys {sample.keys()}.")
if len(matches) > 1 and duplicate_is_error:
raise ValueError(f"Multiple sample keys {sample.keys()} match {pattern}.")
raise ValueError(
f"Multiple sample keys {sample.keys()} match {pattern}.")
value = sample[matches[0]]
result.append(value)
yield tuple(result)
......@@ -484,7 +517,12 @@ def _extract_keys(source, *patterns, duplicate_is_error=True, ignore_missing=Fal
extract_keys = pipelinefilter(_extract_keys)
def _rename_keys(source, *args, keep_unselected=False, must_match=True, duplicate_is_error=True, **kw):
def _rename_keys(source,
*args,
keep_unselected=False,
must_match=True,
duplicate_is_error=True,
**kw):
renamings = [(pattern, output) for output, pattern in args]
renamings += [(pattern, output) for output, pattern in kw.items()]
for sample in source:
......@@ -504,11 +542,15 @@ def _rename_keys(source, *args, keep_unselected=False, must_match=True, duplicat
continue
if new_name in new_sample:
if duplicate_is_error:
raise ValueError(f"Duplicate value in sample {sample.keys()} after rename.")
raise ValueError(
f"Duplicate value in sample {sample.keys()} after rename."
)
continue
new_sample[new_name] = value
if must_match and not all(matched.values()):
raise ValueError(f"Not all patterns ({matched}) matched sample keys ({sample.keys()}).")
raise ValueError(
f"Not all patterns ({matched}) matched sample keys ({sample.keys()})."
)
yield new_sample
......@@ -541,18 +583,18 @@ def find_decoder(decoders, path):
if fname.startswith("__"):
return lambda x: x
for pattern, fun in decoders[::-1]:
if fnmatch(fname.lower(), pattern) or fnmatch("." + fname.lower(), pattern):
if fnmatch(fname.lower(), pattern) or fnmatch("." + fname.lower(),
pattern):
return fun
return None
def _xdecode(
source,
*args,
must_decode=True,
defaults=default_decoders,
**kw,
):
source,
*args,
must_decode=True,
defaults=default_decoders,
**kw, ):
decoders = list(defaults) + list(args)
decoders += [("*." + k, v) for k, v in kw.items()]
for sample in source:
......@@ -575,18 +617,18 @@ def _xdecode(
new_sample[path] = value
yield new_sample
xdecode = pipelinefilter(_xdecode)
xdecode = pipelinefilter(_xdecode)
def _audio_data_filter(source,
frame_shift=10,
max_length=10240,
min_length=10,
token_max_length=200,
token_min_length=1,
min_output_input_ratio=0.0005,
max_output_input_ratio=1):
frame_shift=10,
max_length=10240,
min_length=10,
token_max_length=200,
token_min_length=1,
min_output_input_ratio=0.0005,
max_output_input_ratio=1):
""" Filter sample according to feature and label length
Inplace operation.
......@@ -613,7 +655,8 @@ def _audio_data_filter(source,
assert 'wav' in sample
assert 'label' in sample
# sample['wav'] is paddle.Tensor, we have 100 frames every second (default)
num_frames = sample['wav'].shape[1] / sample['sample_rate'] * (1000 / frame_shift)
num_frames = sample['wav'].shape[1] / sample['sample_rate'] * (
1000 / frame_shift)
if num_frames < min_length:
continue
if num_frames > max_length:
......@@ -629,13 +672,15 @@ def _audio_data_filter(source,
continue
yield sample
audio_data_filter = pipelinefilter(_audio_data_filter)
def _audio_tokenize(source,
symbol_table,
bpe_model=None,
non_lang_syms=None,
split_with_space=False):
symbol_table,
bpe_model=None,
non_lang_syms=None,
split_with_space=False):
""" Decode text to chars or BPE
Inplace operation
......@@ -693,8 +738,10 @@ def _audio_tokenize(source,
sample['label'] = label
yield sample
audio_tokenize = pipelinefilter(_audio_tokenize)
def _audio_resample(source, resample_rate=16000):
""" Resample data.
Inplace operation.
......@@ -713,18 +760,22 @@ def _audio_resample(source, resample_rate=16000):
waveform = sample['wav']
if sample_rate != resample_rate:
sample['sample_rate'] = resample_rate
sample['wav'] = paddle.to_tensor(backends.soundfile_backend.resample(
waveform.numpy(), src_sr = sample_rate, target_sr = resample_rate
))
sample['wav'] = paddle.to_tensor(
backends.soundfile_backend.resample(
waveform.numpy(),
src_sr=sample_rate,
target_sr=resample_rate))
yield sample
audio_resample = pipelinefilter(_audio_resample)
def _audio_compute_fbank(source,
num_mel_bins=80,
frame_length=25,
frame_shift=10,
dither=0.0):
num_mel_bins=80,
frame_length=25,
frame_shift=10,
dither=0.0):
""" Extract fbank
Args:
......@@ -746,30 +797,33 @@ def _audio_compute_fbank(source,
waveform = sample['wav']
waveform = waveform * (1 << 15)
# Only keep fname, feat, label
mat = kaldi.fbank(waveform,
n_mels=num_mel_bins,
frame_length=frame_length,
frame_shift=frame_shift,
dither=dither,
energy_floor=0.0,
sr=sample_rate)
mat = kaldi.fbank(
waveform,
n_mels=num_mel_bins,
frame_length=frame_length,
frame_shift=frame_shift,
dither=dither,
energy_floor=0.0,
sr=sample_rate)
yield dict(fname=sample['fname'], label=sample['label'], feat=mat)
audio_compute_fbank = pipelinefilter(_audio_compute_fbank)
def _audio_spec_aug(source,
max_w=5,
w_inplace=True,
w_mode="PIL",
max_f=30,
num_f_mask=2,
f_inplace=True,
f_replace_with_zero=False,
max_t=40,
num_t_mask=2,
t_inplace=True,
t_replace_with_zero=False,):
def _audio_spec_aug(
source,
max_w=5,
w_inplace=True,
w_mode="PIL",
max_f=30,
num_f_mask=2,
f_inplace=True,
f_replace_with_zero=False,
max_t=40,
num_t_mask=2,
t_inplace=True,
t_replace_with_zero=False, ):
""" Do spec augmentation
Inplace operation
......@@ -793,12 +847,23 @@ def _audio_spec_aug(source,
for sample in source:
x = sample['feat']
x = x.numpy()
x = time_warp(x, max_time_warp=max_w, inplace = w_inplace, mode= w_mode)
x = freq_mask(x, F = max_f, n_mask = num_f_mask, inplace = f_inplace, replace_with_zero = f_replace_with_zero)
x = time_mask(x, T = max_t, n_mask = num_t_mask, inplace = t_inplace, replace_with_zero = t_replace_with_zero)
x = time_warp(x, max_time_warp=max_w, inplace=w_inplace, mode=w_mode)
x = freq_mask(
x,
F=max_f,
n_mask=num_f_mask,
inplace=f_inplace,
replace_with_zero=f_replace_with_zero)
x = time_mask(
x,
T=max_t,
n_mask=num_t_mask,
inplace=t_inplace,
replace_with_zero=t_replace_with_zero)
sample['feat'] = paddle.to_tensor(x, dtype=paddle.float32)
yield sample
audio_spec_aug = pipelinefilter(_audio_spec_aug)
......@@ -829,8 +894,10 @@ def _sort(source, sort_size=500):
for x in buf:
yield x
sort = pipelinefilter(_sort)
def _batched(source, batch_size=16):
""" Static batch the data by `batch_size`
......@@ -850,8 +917,10 @@ def _batched(source, batch_size=16):
if len(buf) > 0:
yield buf
batched = pipelinefilter(_batched)
def dynamic_batched(source, max_frames_in_batch=12000):
""" Dynamic batch the data until the total frames in batch
reach `max_frames_in_batch`
......@@ -892,8 +961,8 @@ def _audio_padding(source):
"""
for sample in source:
assert isinstance(sample, list)
feats_length = paddle.to_tensor([x['feat'].shape[0] for x in sample],
dtype="int64")
feats_length = paddle.to_tensor(
[x['feat'].shape[0] for x in sample], dtype="int64")
order = paddle.argsort(feats_length, descending=True)
feats_lengths = paddle.to_tensor(
[sample[i]['feat'].shape[0] for i in order], dtype="int64")
......@@ -902,20 +971,20 @@ def _audio_padding(source):
sorted_labels = [
paddle.to_tensor(sample[i]['label'], dtype="int32") for i in order
]
label_lengths = paddle.to_tensor([x.shape[0] for x in sorted_labels],
dtype="int64")
padded_feats = pad_sequence(sorted_feats,
batch_first=True,
padding_value=0)
padding_labels = pad_sequence(sorted_labels,
batch_first=True,
padding_value=-1)
yield (sorted_keys, padded_feats, feats_lengths, padding_labels,
label_lengths = paddle.to_tensor(
[x.shape[0] for x in sorted_labels], dtype="int64")
padded_feats = pad_sequence(
sorted_feats, batch_first=True, padding_value=0)
padding_labels = pad_sequence(
sorted_labels, batch_first=True, padding_value=-1)
yield (sorted_keys, padded_feats, feats_lengths, padding_labels,
label_lengths)
audio_padding = pipelinefilter(_audio_padding)
def _audio_cmvn(source, cmvn_file):
global_cmvn = GlobalCMVN(cmvn_file)
for batch in source:
......@@ -923,13 +992,16 @@ def _audio_cmvn(source, cmvn_file):
padded_feats = padded_feats.numpy()
padded_feats = global_cmvn(padded_feats)
padded_feats = paddle.to_tensor(padded_feats, dtype=paddle.float32)
yield (sorted_keys, padded_feats, feats_lengths, padding_labels,
label_lengths)
yield (sorted_keys, padded_feats, feats_lengths, padding_labels,
label_lengths)
audio_cmvn = pipelinefilter(_audio_cmvn)
def _placeholder(source):
for data in source:
yield data
placeholder = pipelinefilter(_placeholder)
paddlespeech/audio/streamdata/handlers.py
浏览文件 @ 598eb1a5
......@@ -3,7 +3,6 @@
# This file is part of the WebDataset library.
# See the LICENSE file for licensing terms (BSD-style).
#
"""Pluggable exception handlers.
These are functions that take an exception as an argument and then return...
......@@ -14,8 +13,8 @@ These are functions that take an exception as an argument and then return...
They are used as handler= arguments in much of the library.
"""
import time, warnings
import time
import warnings
def reraise_exception(exn):
......
paddlespeech/audio/streamdata/mix.py
浏览文件 @ 598eb1a5
......@@ -5,17 +5,12 @@
# See the LICENSE file for licensing terms (BSD-style).
# Modified from https://github.com/webdataset/webdataset
#
"""Classes for mixing samples from multiple sources."""
import itertools, os, random, time, sys
from functools import reduce, wraps
import random
import numpy as np
from . import autodecode, utils
from .paddle_utils import PaddleTensor, IterableDataset
from .utils import PipelineStage
from .paddle_utils import IterableDataset
def round_robin_shortest(*sources):
......
paddlespeech/audio/streamdata/paddle_utils.py
浏览文件 @ 598eb1a5
......@@ -5,12 +5,11 @@
# See the LICENSE file for licensing terms (BSD-style).
# Modified from https://github.com/webdataset/webdataset
#
"""Mock implementations of paddle interfaces when paddle is not available."""
try:
from paddle.io import DataLoader, IterableDataset
from paddle.io import DataLoader
from paddle.io import IterableDataset
except ModuleNotFoundError:
class IterableDataset:
......@@ -22,12 +21,3 @@ except ModuleNotFoundError:
"""Empty implementation of DataLoader when paddle is not available."""
pass
try:
from paddle import Tensor as PaddleTensor
except ModuleNotFoundError:
class TorchTensor:
"""Empty implementation of PaddleTensor when paddle is not available."""
pass
paddlespeech/audio/streamdata/pipeline.py
浏览文件 @ 598eb1a5
......@@ -3,15 +3,12 @@
# See the LICENSE file for licensing terms (BSD-style).
# Modified from https://github.com/webdataset/webdataset
#%%
import copy, os, random, sys, time
from dataclasses import dataclass
import copy
import sys
from itertools import islice
from typing import List
import braceexpand, yaml
from .handlers import reraise_exception
from .paddle_utils import DataLoader, IterableDataset
from .paddle_utils import DataLoader
from .paddle_utils import IterableDataset
from .utils import PipelineStage
......@@ -22,8 +19,7 @@ def add_length_method(obj):
Combined = type(
obj.__class__.__name__ + "_Length",
(obj.__class__, IterableDataset),
{"__len__": length},
)
{"__len__": length}, )
obj.__class__ = Combined
return obj
......
paddlespeech/audio/transform/spec_augment.py
浏览文件 @ 598eb1a5
......@@ -14,6 +14,7 @@
# Modified from espnet(https://github.com/espnet/espnet)
"""Spec Augment module for preprocessing i.e., data augmentation"""
import random
import numpy
from PIL import Image
......
paddlespeech/cli/asr/infer.py
浏览文件 @ 598eb1a5
此差异已折叠。
paddlespeech/cli/executor.py
浏览文件 @ 598eb1a5
此差异已折叠。
paddlespeech/cli/vector/infer.py
浏览文件 @ 598eb1a5
此差异已折叠。
paddlespeech/s2t/__init__.py
浏览文件 @ 598eb1a5
此差异已折叠。
paddlespeech/s2t/exps/u2/model.py
浏览文件 @ 598eb1a5
此差异已折叠。
paddlespeech/s2t/io/dataloader.py
浏览文件 @ 598eb1a5
此差异已折叠。
paddlespeech/s2t/models/u2/u2.py
浏览文件 @ 598eb1a5
此差异已折叠。
paddlespeech/s2t/modules/align.py
浏览文件 @ 598eb1a5
此差异已折叠。
paddlespeech/s2t/modules/encoder.py
浏览文件 @ 598eb1a5
此差异已折叠。
paddlespeech/t2s/exps/fastspeech2/vc2_infer.py 0 → 100644
浏览文件 @ 598eb1a5
此差异已折叠。
paddlespeech/t2s/exps/syn_utils.py
浏览文件 @ 598eb1a5
此差异已折叠。
paddlespeech/t2s/exps/vits/__init__.py 0 → 100644
浏览文件 @ 598eb1a5
此差异已折叠。
paddlespeech/t2s/exps/vits/train.py
浏览文件 @ 598eb1a5
此差异已折叠。
paddlespeech/t2s/exps/vits/voice_cloning.py 0 → 100644
浏览文件 @ 598eb1a5
此差异已折叠。
setup.py
浏览文件 @ 598eb1a5
此差异已折叠。
tests/test_tipc/prepare.sh
浏览文件 @ 598eb1a5
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或