Skip to content

D
DeepSpeech

PaddlePaddle / DeepSpeech
大约 1 年 前同步成功

通知 207
Star 8425
Fork 1598
D
DeepSpeech

体验新版 GitCode，发现更多精彩内容 >>

提交 c4035f8c 编写于 作者: 小湉湉's avatar 小湉湉
浏览文件
操作

Merge branch 'develop' of github.com:PaddlePaddle/PaddleSpeech into add_aishell3_hifigan

上级 8d919781 62cee69c
展开全部 隐藏空白更改
内联 并排
Showing with 17925 addition and 2100 deletion
.gitignore
浏览文件 @ c4035f8c
......@@ -14,6 +14,7 @@
*.whl
*.egg-info
build
*output/
docs/build/
docs/topic/ctc/warp-ctc/
......@@ -33,6 +34,4 @@ tools/activate_python.sh
tools/miniconda.sh
tools/CRF++-0.58/
speechx/fc_patch/
*output/
speechx/fc_patch/
\ No newline at end of file
.pre-commit-config.yaml
浏览文件 @ c4035f8c
......@@ -50,13 +50,13 @@ repos:
entry: bash .pre-commit-hooks/clang-format.hook -i
language: system
files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|cuh|proto)$
exclude: (?=speechx/speechx/kaldi).*(\.cpp|\.cc|\.h|\.py)$
exclude: (?=speechx/speechx/kaldi|speechx/patch).*(\.cpp|\.cc|\.h|\.py)$
- id: copyright_checker
name: copyright_checker
entry: python .pre-commit-hooks/copyright-check.hook
language: system
files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto|py)$
exclude: (?=third_party|pypinyin|speechx/speechx/kaldi).*(\.cpp|\.cc|\.h|\.py)$
exclude: (?=third_party|pypinyin|speechx/speechx/kaldi|speechx/patch).*(\.cpp|\.cc|\.h|\.py)$
- repo: https://github.com/asottile/reorder_python_imports
rev: v2.4.0
hooks:
......
CHANGELOG.md
浏览文件 @ c4035f8c
# Changelog
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:
......
README.md
浏览文件 @ c4035f8c
......@@ -178,7 +178,7 @@ Via the easy-to-use, efficient, flexible and scalable implementation, our vision
<!---
2021.12.14: We would like to have an online courses to introduce basics and research of speech, as well as code practice with `paddlespeech`. Please pay attention to our [Calendar](https://www.paddlepaddle.org.cn/live).
--->
- 🤗 2021.12.14: Our PaddleSpeech [ASR](https://huggingface.co/spaces/KPatrick/PaddleSpeechASR) and [TTS](https://huggingface.co/spaces/akhaliq/paddlespeech) Demos on Hugging Face Spaces are available!
- 🤗 2021.12.14: Our PaddleSpeech [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: PaddleSpeech CLI is available for Audio Classification, Automatic Speech Recognition, Speech Translation (English to Chinese) and Text-to-Speech.
### Community
......@@ -207,6 +207,7 @@ paddlespeech cls --input input.wav
```shell
paddlespeech asr --lang zh --input input_16k.wav
```
- web demo for Automatic Speech Recognition is integrated to [Huggingface Spaces](https://huggingface.co/spaces) with [Gradio](https://github.com/gradio-app/gradio). See Demo: [ASR Demo](https://huggingface.co/spaces/KPatrick/PaddleSpeechASR)
**Speech Translation** (English to Chinese)
(not support for Mac and Windows now)
......@@ -218,7 +219,7 @@ paddlespeech st --input input_16k.wav
```shell
paddlespeech tts --input "你好，欢迎使用飞桨深度学习框架！" --output output.wav
```
- web demo for Text to Speech is integrated to [Huggingface Spaces](https://huggingface.co/spaces) with [Gradio](https://github.com/gradio-app/gradio). See Demo: [TTS Demo](https://huggingface.co/spaces/akhaliq/paddlespeech)
- web demo for Text to Speech is integrated to [Huggingface Spaces](https://huggingface.co/spaces) with [Gradio](https://github.com/gradio-app/gradio). See Demo: [TTS Demo](https://huggingface.co/spaces/KPatrick/PaddleSpeechTTS)
**Text Postprocessing**
- Punctuation Restoration
......@@ -397,9 +398,9 @@ PaddleSpeech supports a series of most popular models. They are summarized in [r
</tr>
<tr>
<td >HiFiGAN</td>
<td >CSMSC</td>
<td >LJSpeech / VCTK / CSMSC / AISHELL-3</td>
<td>
<a href = "./examples/csmsc/voc5">HiFiGAN-csmsc</a>
<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>
......@@ -573,7 +574,6 @@ You are warmly welcome to submit questions in [discussions](https://github.com/P
- 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 [AK391](https://github.com/AK391) for TTS web demo on Huggingface Spaces using Gradio.
- 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.
- Many thanks to [745165806](https://github.com/745165806)/[PaddleSpeechTask](https://github.com/745165806/PaddleSpeechTask) for contributing Punctuation Restoration model.
......
README_cn.md
浏览文件 @ c4035f8c
......@@ -392,9 +392,9 @@ PaddleSpeech 的 **语音合成** 主要包含三个模块：文本前端、声
</tr>
<tr>
<td >HiFiGAN</td>
<td >CSMSC</td>
<td >LJSpeech / VCTK / CSMSC / AISHELL-3</td>
<td>
<a href = "./examples/csmsc/voc5">HiFiGAN-csmsc</a>
<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>
......
demos/speech_recognition/README.md
浏览文件 @ c4035f8c
......@@ -84,5 +84,8 @@ Here is a list of pretrained models released by PaddleSpeech that can be used by
| Model | Language | Sample Rate
| :--- | :---: | :---: |
| conformer_wenetspeech| zh| 16000
| transformer_librispeech| en| 16000
| conformer_wenetspeech| zh| 16k
| transformer_librispeech| en| 16k
| deepspeech2offline_aishell| zh| 16k
| deepspeech2online_aishell | zh | 16k
|deepspeech2offline_librispeech|en| 16k
demos/speech_recognition/README_cn.md
浏览文件 @ c4035f8c
......@@ -81,5 +81,8 @@ wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav https://paddlespee
| 模型 | 语言 | 采样率
| :--- | :---: | :---: |
| conformer_wenetspeech| zh| 16000
| transformer_librispeech| en| 16000
| conformer_wenetspeech | zh | 16k
| transformer_librispeech | en | 16k
| deepspeech2offline_aishell| zh| 16k
| deepspeech2online_aishell | zh | 16k
| deepspeech2offline_librispeech | en | 16k
demos/speech_server/README.md
浏览文件 @ c4035f8c
......@@ -10,21 +10,15 @@ This demo is an implementation of starting the voice service and accessing the s
### 1. Installation
see [installation](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/docs/source/install.md).
You can choose one way from easy, meduim and hard to install paddlespeech.
It is recommended to use **paddlepaddle 2.2.1** or above.
You can choose one way from meduim and hard to install paddlespeech.
### 2. Prepare config File
The configuration file contains the service-related configuration files and the model configuration related to the voice tasks contained in the service. They are all under the `conf` folder.
The configuration file can be found in `conf/application.yaml` .
Among them, `engine_list` indicates the speech engine that will be included in the service to be started, in the format of <speech task>_<engine type>.
At present, the speech tasks integrated by the service include: asr (speech recognition) and tts (speech synthesis).
Currently the engine type supports two forms: python and inference (Paddle Inference)
**Note: The configuration of `engine_backend` in `application.yaml` represents all speech tasks included in the started service. **
If the service you want to start contains only a certain speech task, then you need to comment out the speech tasks that do not need to be included. For example, if you only want to use the speech recognition (ASR) service, then you can comment out the speech synthesis (TTS) service, as in the following example:
```bash
engine_backend:
asr: 'conf/asr/asr.yaml'
#tts: 'conf/tts/tts.yaml'
```
**Note: The configuration file of `engine_backend` in `application.yaml` needs to match the configuration type of `engine_type`. **
When the configuration file of `engine_backend` is `XXX.yaml`, the configuration type of `engine_type` needs to be set to `python`; when the configuration file of `engine_backend` is `XXX_pd.yaml`, the configuration of `engine_type` needs to be set type is `inference`;
The input of ASR client demo should be a WAV file(`.wav`), and the sample rate must be the same as the model.
......@@ -116,21 +110,22 @@ wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav https://paddlespee
- Python API
```python
from paddlespeech.server.bin.paddlespeech_client import ASRClientExecutor
import json
asrclient_executor = ASRClientExecutor()
asrclient_executor(
res = asrclient_executor(
input="./zh.wav",
server_ip="127.0.0.1",
port=8090,
sample_rate=16000,
lang="zh_cn",
audio_format="wav")
print(res.json())
```
Output:
```bash
{'success': True, 'code': 200, 'message': {'description': 'success'}, 'result': {'transcription': '我认为跑步最重要的就是给我带来了身体健康'}}
time cost 0.604353 s.
```
### 5. TTS Client Usage
......@@ -152,7 +147,7 @@ wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav https://paddlespee
- `speed`: Audio speed, the value should be set between 0 and 3. Default: 1.0
- `volume`: Audio volume, the value should be set between 0 and 3. Default: 1.0
- `sample_rate`: Sampling rate, choice: [0, 8000, 16000], the default is the same as the model. Default: 0
- `output`: Output wave filepath. Default: `output.wav`.
- `output`: Output wave filepath. Default: None, which means not to save the audio to the local.
Output:
```bash
......@@ -166,9 +161,10 @@ wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav https://paddlespee
- Python API
```python
from paddlespeech.server.bin.paddlespeech_client import TTSClientExecutor
import json
ttsclient_executor = TTSClientExecutor()
ttsclient_executor(
res = ttsclient_executor(
input="您好，欢迎使用百度飞桨语音合成服务。",
server_ip="127.0.0.1",
port=8090,
......@@ -177,6 +173,11 @@ wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav https://paddlespee
volume=1.0,
sample_rate=0,
output="./output.wav")
response_dict = res.json()
print(response_dict["message"])
print("Save synthesized audio successfully on %s." % (response_dict['result']['save_path']))
print("Audio duration: %f s." %(response_dict['result']['duration']))
```
Output:
......@@ -184,7 +185,52 @@ wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav https://paddlespee
{'description': 'success.'}
Save synthesized audio successfully on ./output.wav.
Audio duration: 3.612500 s.
Response time: 0.388317 s.
```
### 6. CLS Client Usage
**Note:** The response time will be slightly longer when using the client for the first time
- Command Line (Recommended)
```
paddlespeech_client cls --server_ip 127.0.0.1 --port 8090 --input ./zh.wav
```
Usage:
```bash
paddlespeech_client cls --help
```
Arguments:
- `server_ip`: server ip. Default: 127.0.0.1
- `port`: server port. Default: 8090
- `input`(required): Audio file to be classified.
- `topk`: topk scores of classification result.
Output:
```bash
[2022-03-09 20:44:39,974] [ INFO] - {'success': True, 'code': 200, 'message': {'description': 'success'}, 'result': {'topk': 1, 'results': [{'class_name': 'Speech', 'prob': 0.9027184844017029}]}}
[2022-03-09 20:44:39,975] [ INFO] - Response time 0.104360 s.
```
- Python API
```python
from paddlespeech.server.bin.paddlespeech_client import CLSClientExecutor
import json
clsclient_executor = CLSClientExecutor()
res = clsclient_executor(
input="./zh.wav",
server_ip="127.0.0.1",
port=8090,
topk=1)
print(res.json())
```
Output:
```bash
{'success': True, 'code': 200, 'message': {'description': 'success'}, 'result': {'topk': 1, 'results': [{'class_name': 'Speech', 'prob': 0.9027184844017029}]}}
```
......@@ -195,3 +241,6 @@ Get all models supported by the ASR service via `paddlespeech_server stats --tas
### TTS model
Get all models supported by the TTS service via `paddlespeech_server stats --task tts`, where static models can be used for paddle inference inference.
### CLS model
Get all models supported by the CLS service via `paddlespeech_server stats --task cls`, where static models can be used for paddle inference inference.
demos/speech_server/README_cn.md
浏览文件 @ c4035f8c
......@@ -10,19 +10,16 @@
### 1. 安装
请看 [安装文档](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/docs/source/install.md).
你可以从 easy，medium，hard 三中方式中选择一种方式安装 PaddleSpeech。
推荐使用 **paddlepaddle 2.2.1** 或以上版本。
你可以从 medium，hard 三中方式中选择一种方式安装 PaddleSpeech。
### 2. 准备配置文件
配置文件包含服务相关的配置文件和服务中包含的语音任务相关的模型配置。 它们都在 `conf` 文件夹下。
**注意：`application.yaml` 中 `engine_backend` 的配置表示启动的服务中包含的所有语音任务。**
如果你想启动的服务中只包含某项语音任务，那么你需要注释掉不需要包含的语音任务。例如你只想使用语音识别（ASR）服务，那么你可以将语音合成（TTS）服务注释掉，如下示例：
```bash
engine_backend:
asr: 'conf/asr/asr.yaml'
#tts: 'conf/tts/tts.yaml'
```
**注意：`application.yaml` 中 `engine_backend` 的配置文件需要和 `engine_type` 的配置类型匹配。**
当`engine_backend` 的配置文件为`XXX.yaml`时，需要设置`engine_type`的配置类型为`python`;当`engine_backend` 的配置文件为`XXX_pd.yaml`时，需要设置`engine_type`的配置类型为`inference`;
配置文件可参见 `conf/application.yaml` 。
其中，`engine_list`表示即将启动的服务将会包含的语音引擎，格式为 <语音任务>_<引擎类型>。
目前服务集成的语音任务有： asr(语音识别)、tts(语音合成)。
目前引擎类型支持两种形式：python 及 inference (Paddle Inference)
这个 ASR client 的输入应该是一个 WAV 文件（`.wav`），并且采样率必须与模型的采样率相同。
......@@ -83,8 +80,8 @@ wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav https://paddlespee
```
### 4. ASR客户端使用方法
**注意：**初次使用客户端时响应时间会略长
### 4. ASR 客户端使用方法
**注意：** 初次使用客户端时响应时间会略长
- 命令行 (推荐使用)
```
paddlespeech_client asr --server_ip 127.0.0.1 --port 8090 --input ./zh.wav
......@@ -114,29 +111,32 @@ wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav https://paddlespee
- Python API
```python
from paddlespeech.server.bin.paddlespeech_client import ASRClientExecutor
import json
asrclient_executor = ASRClientExecutor()
asrclient_executor(
res = asrclient_executor(
input="./zh.wav",
server_ip="127.0.0.1",
port=8090,
sample_rate=16000,
lang="zh_cn",
audio_format="wav")
print(res.json())
```
输出:
```bash
{'success': True, 'code': 200, 'message': {'description': 'success'}, 'result': {'transcription': '我认为跑步最重要的就是给我带来了身体健康'}}
time cost 0.604353 s.
```
### 5. TTS客户端使用方法
**注意：**初次使用客户端时响应时间会略长
```bash
paddlespeech_client tts --server_ip 127.0.0.1 --port 8090 --input "您好，欢迎使用百度飞桨语音合成服务。" --output output.wav
```
### 5. TTS 客户端使用方法
**注意：** 初次使用客户端时响应时间会略长
- 命令行 (推荐使用)
```bash
paddlespeech_client tts --server_ip 127.0.0.1 --port 8090 --input "您好，欢迎使用百度飞桨语音合成服务。" --output output.wav
```
使用帮助:
```bash
......@@ -151,7 +151,7 @@ wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav https://paddlespee
- `speed`: 音频速度，该值应设置在 0 到 3 之间。 默认值：1.0
- `volume`: 音频音量，该值应设置在 0 到 3 之间。 默认值： 1.0
- `sample_rate`: 采样率，可选 [0, 8000, 16000]，默认与模型相同。 默认值：0
- `output`: 输出音频的路径， 默认值：output.wav。
- `output`: 输出音频的路径， 默认值：None，表示不保存音频到本地。
输出:
```bash
......@@ -164,9 +164,10 @@ wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav https://paddlespee
- Python API
```python
from paddlespeech.server.bin.paddlespeech_client import TTSClientExecutor
import json
ttsclient_executor = TTSClientExecutor()
ttsclient_executor(
res = ttsclient_executor(
input="您好，欢迎使用百度飞桨语音合成服务。",
server_ip="127.0.0.1",
port=8090,
......@@ -175,6 +176,11 @@ wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav https://paddlespee
volume=1.0,
sample_rate=0,
output="./output.wav")
response_dict = res.json()
print(response_dict["message"])
print("Save synthesized audio successfully on %s." % (response_dict['result']['save_path']))
print("Audio duration: %f s." %(response_dict['result']['duration']))
```
输出:
......@@ -182,13 +188,63 @@ wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav https://paddlespee
{'description': 'success.'}
Save synthesized audio successfully on ./output.wav.
Audio duration: 3.612500 s.
Response time: 0.388317 s.
```
### 5. CLS 客户端使用方法
**注意：** 初次使用客户端时响应时间会略长
- 命令行 (推荐使用)
```
paddlespeech_client cls --server_ip 127.0.0.1 --port 8090 --input ./zh.wav
```
使用帮助:
```bash
paddlespeech_client cls --help
```
参数:
- `server_ip`: 服务端ip地址，默认: 127.0.0.1。
- `port`: 服务端口，默认: 8090。
- `input`(必须输入): 用于分类的音频文件。
- `topk`: 分类结果的topk。
输出:
```bash
[2022-03-09 20:44:39,974] [ INFO] - {'success': True, 'code': 200, 'message': {'description': 'success'}, 'result': {'topk': 1, 'results': [{'class_name': 'Speech', 'prob': 0.9027184844017029}]}}
[2022-03-09 20:44:39,975] [ INFO] - Response time 0.104360 s.
```
- Python API
```python
from paddlespeech.server.bin.paddlespeech_client import CLSClientExecutor
import json
clsclient_executor = CLSClientExecutor()
res = clsclient_executor(
input="./zh.wav",
server_ip="127.0.0.1",
port=8090,
topk=1)
print(res.json())
```
输出:
```bash
{'success': True, 'code': 200, 'message': {'description': 'success'}, 'result': {'topk': 1, 'results': [{'class_name': 'Speech', 'prob': 0.9027184844017029}]}}
```
## 服务支持的模型
### ASR支持的模型
通过 `paddlespeech_server stats --task asr` 获取ASR服务支持的所有模型，其中静态模型可用于 paddle inference 推理。
### TTS支持的模型
通过 `paddlespeech_server stats --task tts` 获取TTS服务支持的所有模型，其中静态模型可用于 paddle inference 推理。
### CLS支持的模型
通过 `paddlespeech_server stats --task cls` 获取CLS服务支持的所有模型，其中静态模型可用于 paddle inference 推理。
demos/speech_server/cls_client.sh 0 → 100644
浏览文件 @ c4035f8c
#!/bin/bash
wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav https://paddlespeech.bj.bcebos.com/PaddleAudio/en.wav
paddlespeech_client cls --server_ip 127.0.0.1 --port 8090 --input ./zh.wav --topk 1
demos/speech_server/conf/application.yaml
浏览文件 @ c4035f8c
# This is the parameter configuration file for PaddleSpeech Serving.
##################################################################
# SERVER SETTING #
##################################################################
host: '127.0.0.1'
#################################################################################
# SERVER SETTING #
#################################################################################
host: 127.0.0.1
port: 8090
##################################################################
# CONFIG FILE #
##################################################################
# add engine backend type (Options: asr, tts) and config file here.
# Adding a speech task to engine_backend means starting the service.
engine_backend:
asr: 'conf/asr/asr.yaml'
tts: 'conf/tts/tts.yaml'
# The engine_type of speech task needs to keep the same type as the config file of speech task.
# E.g: The engine_type of asr is 'python', the engine_backend of asr is 'XX/asr.yaml'
# E.g: The engine_type of asr is 'inference', the engine_backend of asr is 'XX/asr_pd.yaml'
#
# add engine type (Options: python, inference)
engine_type:
asr: 'python'
tts: 'python'
# The task format in the engin_list is: <speech task>_<engine type>
# task choices = ['asr_python', 'asr_inference', 'tts_python', 'tts_inference']
engine_list: ['asr_python', 'tts_python', 'cls_python']
#################################################################################
# ENGINE CONFIG #
#################################################################################
################################### ASR #########################################
################### speech task: asr; engine_type: python #######################
asr_python:
model: 'conformer_wenetspeech'
lang: 'zh'
sample_rate: 16000
cfg_path: # [optional]
ckpt_path: # [optional]
decode_method: 'attention_rescoring'
force_yes: True
device: # set 'gpu:id' or 'cpu'
################### speech task: asr; engine_type: inference #######################
asr_inference:
# model_type choices=['deepspeech2offline_aishell']
model_type: 'deepspeech2offline_aishell'
am_model: # the pdmodel file of am static model [optional]
am_params: # the pdiparams file of am static model [optional]
lang: 'zh'
sample_rate: 16000
cfg_path:
decode_method:
force_yes: True
am_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
################################### TTS #########################################
################### speech task: tts; engine_type: python #######################
tts_python:
# am (acoustic model) choices=['speedyspeech_csmsc', 'fastspeech2_csmsc',
# 'fastspeech2_ljspeech', 'fastspeech2_aishell3',
# 'fastspeech2_vctk']
am: 'fastspeech2_csmsc'
am_config:
am_ckpt:
am_stat:
phones_dict:
tones_dict:
speaker_dict:
spk_id: 0
# voc (vocoder) choices=['pwgan_csmsc', 'pwgan_ljspeech', 'pwgan_aishell3',
# 'pwgan_vctk', 'mb_melgan_csmsc']
voc: 'pwgan_csmsc'
voc_config:
voc_ckpt:
voc_stat:
# others
lang: 'zh'
device: # set 'gpu:id' or 'cpu'
################### speech task: tts; engine_type: inference #######################
tts_inference:
# am (acoustic model) choices=['speedyspeech_csmsc', 'fastspeech2_csmsc']
am: 'fastspeech2_csmsc'
am_model: # the pdmodel file of your am static model (XX.pdmodel)
am_params: # the pdiparams file of your am static model (XX.pdipparams)
am_sample_rate: 24000
phones_dict:
tones_dict:
speaker_dict:
spk_id: 0
am_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
# voc (vocoder) choices=['pwgan_csmsc', 'mb_melgan_csmsc','hifigan_csmsc']
voc: 'pwgan_csmsc'
voc_model: # the pdmodel file of your vocoder static model (XX.pdmodel)
voc_params: # the pdiparams file of your vocoder static model (XX.pdipparams)
voc_sample_rate: 24000
voc_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
# others
lang: 'zh'
################################### CLS #########################################
################### speech task: cls; engine_type: python #######################
cls_python:
# model choices=['panns_cnn14', 'panns_cnn10', 'panns_cnn6']
model: 'panns_cnn14'
cfg_path: # [optional] Config of cls task.
ckpt_path: # [optional] Checkpoint file of model.
label_file: # [optional] Label file of cls task.
device: # set 'gpu:id' or 'cpu'
################### speech task: cls; engine_type: inference #######################
cls_inference:
# model_type choices=['panns_cnn14', 'panns_cnn10', 'panns_cnn6']
model_type: 'panns_cnn14'
cfg_path:
model_path: # the pdmodel file of am static model [optional]
params_path: # the pdiparams file of am static model [optional]
label_file: # [optional] Label file of cls task.
predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
demos/speech_server/conf/asr/asr.yaml 已删除 100644 → 0
浏览文件 @ 8d919781
model: 'conformer_wenetspeech'
lang: 'zh'
sample_rate: 16000
cfg_path: # [optional]
ckpt_path: # [optional]
decode_method: 'attention_rescoring'
force_yes: True
device: # set 'gpu:id' or 'cpu'
demos/speech_server/conf/asr/asr_pd.yaml 已删除 100644 → 0
浏览文件 @ 8d919781
# This is the parameter configuration file for ASR server.
# These are the static models that support paddle inference.
##################################################################
# ACOUSTIC MODEL SETTING #
# am choices=['deepspeech2offline_aishell'] TODO
##################################################################
model_type: 'deepspeech2offline_aishell'
am_model: # the pdmodel file of am static model [optional]
am_params: # the pdiparams file of am static model [optional]
lang: 'zh'
sample_rate: 16000
cfg_path:
decode_method:
force_yes: True
am_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
##################################################################
# OTHERS #
##################################################################
demos/speech_server/conf/tts/tts.yaml 已删除 100644 → 0
浏览文件 @ 8d919781
# This is the parameter configuration file for TTS server.
##################################################################
# ACOUSTIC MODEL SETTING #
# am choices=['speedyspeech_csmsc', 'fastspeech2_csmsc',
# 'fastspeech2_ljspeech', 'fastspeech2_aishell3',
# 'fastspeech2_vctk']
##################################################################
am: 'fastspeech2_csmsc'
am_config:
am_ckpt:
am_stat:
phones_dict:
tones_dict:
speaker_dict:
spk_id: 0
##################################################################
# VOCODER SETTING #
# voc choices=['pwgan_csmsc', 'pwgan_ljspeech', 'pwgan_aishell3',
# 'pwgan_vctk', 'mb_melgan_csmsc']
##################################################################
voc: 'pwgan_csmsc'
voc_config:
voc_ckpt:
voc_stat:
##################################################################
# OTHERS #
##################################################################
lang: 'zh'
device: # set 'gpu:id' or 'cpu'
demos/speech_server/conf/tts/tts_pd.yaml 已删除 100644 → 0
浏览文件 @ 8d919781
# This is the parameter configuration file for TTS server.
# These are the static models that support paddle inference.
##################################################################
# ACOUSTIC MODEL SETTING #
# am choices=['speedyspeech_csmsc', 'fastspeech2_csmsc']
##################################################################
am: 'fastspeech2_csmsc'
am_model: # the pdmodel file of your am static model (XX.pdmodel)
am_params: # the pdiparams file of your am static model (XX.pdipparams)
am_sample_rate: 24000
phones_dict:
tones_dict:
speaker_dict:
spk_id: 0
am_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
##################################################################
# VOCODER SETTING #
# voc choices=['pwgan_csmsc', 'mb_melgan_csmsc','hifigan_csmsc']
##################################################################
voc: 'pwgan_csmsc'
voc_model: # the pdmodel file of your vocoder static model (XX.pdmodel)
voc_params: # the pdiparams file of your vocoder static model (XX.pdipparams)
voc_sample_rate: 24000
voc_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
##################################################################
# OTHERS #
##################################################################
lang: 'zh'
demos/speech_server/server.sh
浏览文件 @ c4035f8c
#!/bin/bash
paddlespeech_server start --config_file ./conf/application.yaml
\ No newline at end of file
paddlespeech_server start --config_file ./conf/application.yaml
docs/source/reference.md
浏览文件 @ c4035f8c
......@@ -35,3 +35,7 @@ We borrowed a lot of code from these repos to build `model` and `engine`, thanks
* [librosa](https://github.com/librosa/librosa/blob/main/LICENSE.md)
- ISC License
- Audio feature
* [ThreadPool](https://github.com/progschj/ThreadPool/blob/master/COPYING)
- zlib License
- ThreadPool
docs/source/released_model.md
浏览文件 @ c4035f8c
......@@ -59,7 +59,7 @@ WaveRNN | CSMSC |[WaveRNN-csmsc](https://github.com/PaddlePaddle/PaddleSpeech/tr
### Voice Cloning
Model Type | Dataset| Example Link | Pretrained Models
:-------------:| :------------:| :-----: | :-----:
:-------------:| :------------:| :-----: | :-----: |
GE2E| AISHELL-3, etc. |[ge2e](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/ge2e)|[ge2e_ckpt_0.3.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/ge2e/ge2e_ckpt_0.3.zip)
GE2E + Tactron2| AISHELL-3 |[ge2e-tactron2-aishell3](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/aishell3/vc0)|[tacotron2_aishell3_ckpt_vc0_0.2.0.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/tacotron2/tacotron2_aishell3_ckpt_vc0_0.2.0.zip)
GE2E + FastSpeech2 | AISHELL-3 |[ge2e-fastspeech2-aishell3](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/aishell3/vc1)|[fastspeech2_nosil_aishell3_vc1_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_nosil_aishell3_vc1_ckpt_0.5.zip)
......@@ -67,9 +67,9 @@ GE2E + FastSpeech2 | AISHELL-3 |[ge2e-fastspeech2-aishell3](https://github.com/
## Audio Classification Models
Model Type | Dataset| Example Link | Pretrained Models
:-------------:| :------------:| :-----: | :-----:
PANN | Audioset| [audioset_tagging_cnn](https://github.com/qiuqiangkong/audioset_tagging_cnn) | [panns_cnn6.pdparams](https://bj.bcebos.com/paddleaudio/models/panns_cnn6.pdparams), [panns_cnn10.pdparams](https://bj.bcebos.com/paddleaudio/models/panns_cnn10.pdparams), [panns_cnn14.pdparams](https://bj.bcebos.com/paddleaudio/models/panns_cnn14.pdparams)
Model Type | Dataset| Example Link | Pretrained Models | Static Models
:-------------:| :------------:| :-----: | :-----: | :-----:
PANN | Audioset| [audioset_tagging_cnn](https://github.com/qiuqiangkong/audioset_tagging_cnn) | [panns_cnn6.pdparams](https://bj.bcebos.com/paddleaudio/models/panns_cnn6.pdparams), [panns_cnn10.pdparams](https://bj.bcebos.com/paddleaudio/models/panns_cnn10.pdparams), [panns_cnn14.pdparams](https://bj.bcebos.com/paddleaudio/models/panns_cnn14.pdparams) | [panns_cnn6_static.tar.gz](https://paddlespeech.bj.bcebos.com/cls/inference_model/panns_cnn6_static.tar.gz)(18M), [panns_cnn10_static.tar.gz](https://paddlespeech.bj.bcebos.com/cls/inference_model/panns_cnn10_static.tar.gz)(19M), [panns_cnn14_static.tar.gz](https://paddlespeech.bj.bcebos.com/cls/inference_model/panns_cnn14_static.tar.gz)(289M)
PANN | ESC-50 |[pann-esc50](../../examples/esc50/cls0)|[esc50_cnn6.tar.gz](https://paddlespeech.bj.bcebos.com/cls/esc50/esc50_cnn6.tar.gz), [esc50_cnn10.tar.gz](https://paddlespeech.bj.bcebos.com/cls/esc50/esc50_cnn10.tar.gz), [esc50_cnn14.tar.gz](https://paddlespeech.bj.bcebos.com/cls/esc50/esc50_cnn14.tar.gz)
## Punctuation Restoration Models
......
examples/aishell3/vc0/local/preprocess.sh
浏览文件 @ c4035f8c
#!/bin/bash
stage=3
stage=0
stop_stage=100
config_path=$1
......
examples/aishell3/voc1/run.sh
浏览文件 @ c4035f8c
......@@ -3,7 +3,7 @@
set -e
source path.sh
gpus=0
gpus=0,1
stage=0
stop_stage=100
......
examples/csmsc/tts0/local/synthesize_e2e.sh
浏览文件 @ c4035f8c
......@@ -7,7 +7,7 @@ ckpt_name=$3
stage=0
stop_stage=0
# TODO: tacotron2 动转静的结果没有静态图的响亮, 可能还是 decode 的时候某个函数动静不对齐
# TODO: tacotron2 动转静的结果没有动态图的响亮, 可能还是 decode 的时候某个函数动静不对齐
# pwgan
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
FLAGS_allocator_strategy=naive_best_fit \
......
examples/csmsc/tts2/local/synthesize.sh
浏览文件 @ c4035f8c
......@@ -14,7 +14,7 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
--am=speedyspeech_csmsc \
--am_config=${config_path} \
--am_ckpt=${train_output_path}/checkpoints/${ckpt_name} \
--am_stat=dump/train/speech_stats.npy \
--am_stat=dump/train/feats_stats.npy \
--voc=pwgan_csmsc \
--voc_config=pwg_baker_ckpt_0.4/pwg_default.yaml \
--voc_ckpt=pwg_baker_ckpt_0.4/pwg_snapshot_iter_400000.pdz \
......@@ -34,7 +34,7 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
--am=speedyspeech_csmsc \
--am_config=${config_path} \
--am_ckpt=${train_output_path}/checkpoints/${ckpt_name} \
--am_stat=dump/train/speech_stats.npy \
--am_stat=dump/train/feats_stats.npy \
--voc=mb_melgan_csmsc \
--voc_config=mb_melgan_csmsc_ckpt_0.1.1/default.yaml \
--voc_ckpt=mb_melgan_csmsc_ckpt_0.1.1/snapshot_iter_1000000.pdz\
......@@ -53,7 +53,7 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
--am=speedyspeech_csmsc \
--am_config=${config_path} \
--am_ckpt=${train_output_path}/checkpoints/${ckpt_name} \
--am_stat=dump/train/speech_stats.npy \
--am_stat=dump/train/feats_stats.npy \
--voc=style_melgan_csmsc \
--voc_config=style_melgan_csmsc_ckpt_0.1.1/default.yaml \
--voc_ckpt=style_melgan_csmsc_ckpt_0.1.1/snapshot_iter_1500000.pdz \
......@@ -73,7 +73,7 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
--am=speedyspeech_csmsc \
--am_config=${config_path} \
--am_ckpt=${train_output_path}/checkpoints/${ckpt_name} \
--am_stat=dump/train/speech_stats.npy \
--am_stat=dump/train/feats_stats.npy \
--voc=hifigan_csmsc \
--voc_config=hifigan_csmsc_ckpt_0.1.1/default.yaml \
--voc_ckpt=hifigan_csmsc_ckpt_0.1.1/snapshot_iter_2500000.pdz \
......@@ -93,7 +93,7 @@ if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
--am=speedyspeech_csmsc \
--am_config=${config_path} \
--am_ckpt=${train_output_path}/checkpoints/${ckpt_name} \
--am_stat=dump/train/speech_stats.npy \
--am_stat=dump/train/feats_stats.npy \
--voc=wavernn_csmsc \
--voc_config=wavernn_csmsc_ckpt_0.2.0/default.yaml \
--voc_ckpt=wavernn_csmsc_ckpt_0.2.0/snapshot_iter_400000.pdz \
......
examples/ljspeech/voc5/README.md 0 → 100644
浏览文件 @ c4035f8c
# HiFiGAN with the LJSpeech-1.1
This example contains code used to train a [HiFiGAN](https://arxiv.org/abs/2010.05646) model with [LJSpeech-1.1](https://keithito.com/LJ-Speech-Dataset/).
## Dataset
### Download and Extract
Download LJSpeech-1.1 from the [official website](https://keithito.com/LJ-Speech-Dataset/).
### Get MFA Result and Extract
We use [MFA](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) results to cut the silence in the edge of audio.
You can download from here [ljspeech_alignment.tar.gz](https://paddlespeech.bj.bcebos.com/MFA/LJSpeech-1.1/ljspeech_alignment.tar.gz), or train your MFA model reference to [mfa example](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/mfa) of our repo.
## Get Started
Assume the path to the dataset is `~/datasets/LJSpeech-1.1`.
Assume the path to the MFA result of LJSpeech-1.1 is `./ljspeech_alignment`.
Run the command below to
1. **source path**.
2. preprocess the dataset.
3. train the model.
4. synthesize wavs.
- synthesize waveform from `metadata.jsonl`.
```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 ${conf_path}
```
When it is done. A `dump` folder is created in the current directory. The structure of the dump folder is listed below.
```text
dump
├── dev
│ ├── norm
│ └── raw
├── test
│ ├── norm
│ └── raw
└── train
├── norm
├── raw
└── feats_stats.npy
```
The dataset is split into 3 parts, namely `train`, `dev`, and `test`, each of which contains a `norm` and `raw` subfolder. The `raw` folder contains the log magnitude of the mel spectrogram of each utterance, while the norm folder contains the normalized spectrogram. The statistics used to normalize the spectrogram are computed from the training set, which is located in `dump/train/feats_stats.npy`.
Also, there is a `metadata.jsonl` in each subfolder. It is a table-like file that contains id and paths to the spectrogram of each utterance.
### Model Training
`./local/train.sh` calls `${BIN_DIR}/train.py`.
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path}
```
Here's the complete help message.
```text
usage: train.py [-h] [--config CONFIG] [--train-metadata TRAIN_METADATA]
[--dev-metadata DEV_METADATA] [--output-dir OUTPUT_DIR]
[--ngpu NGPU] [--batch-size BATCH_SIZE] [--max-iter MAX_ITER]
[--run-benchmark RUN_BENCHMARK]
[--profiler_options PROFILER_OPTIONS]
Train a ParallelWaveGAN model.
optional arguments:
-h, --help show this help message and exit
--config CONFIG config file to overwrite default config.
--train-metadata TRAIN_METADATA
training data.
--dev-metadata DEV_METADATA
dev data.
--output-dir OUTPUT_DIR
output dir.
--ngpu NGPU if ngpu == 0, use cpu.
benchmark:
arguments related to benchmark.
--batch-size BATCH_SIZE
batch size.
--max-iter MAX_ITER train max steps.
--run-benchmark RUN_BENCHMARK
runing benchmark or not, if True, use the --batch-size
and --max-iter.
--profiler_options PROFILER_OPTIONS
The option of profiler, which should be in format
"key1=value1;key2=value2;key3=value3".
```
1. `--config` is a config file in yaml format to overwrite the default config, which can be found at `conf/default.yaml`.
2. `--train-metadata` and `--dev-metadata` should be the metadata file in the normalized subfolder of `train` and `dev` in the `dump` folder.
3. `--output-dir` is the directory to save the results of the experiment. Checkpoints are saved in `checkpoints/` inside this directory.
4. `--ngpu` is the number of gpus to use, if ngpu == 0, use cpu.
### Synthesizing
`./local/synthesize.sh` calls `${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] [--generator-type GENERATOR_TYPE] [--config CONFIG]
[--checkpoint CHECKPOINT] [--test-metadata TEST_METADATA]
[--output-dir OUTPUT_DIR] [--ngpu NGPU]
Synthesize with GANVocoder.
optional arguments:
-h, --help show this help message and exit
--generator-type GENERATOR_TYPE
type of GANVocoder, should in {pwgan, mb_melgan,
style_melgan, } now
--config CONFIG GANVocoder config file.
--checkpoint CHECKPOINT
snapshot to load.
--test-metadata TEST_METADATA
dev data.
--output-dir OUTPUT_DIR
output dir.
--ngpu NGPU if ngpu == 0, use cpu.
```
1. `--config` parallel wavegan config file. You should use the same config with which the model is trained.
2. `--checkpoint` is the checkpoint to load. Pick one of the checkpoints from `checkpoints` inside the training output directory.
3. `--test-metadata` is the metadata of the test dataset. Use the `metadata.jsonl` in the `dev/norm` subfolder from the processed directory.
4. `--output-dir` is the directory to save the synthesized audio files.
5. `--ngpu` is the number of gpus to use, if ngpu == 0, use cpu.
## Pretrained Model
## Acknowledgement
We adapted some code from https://github.com/kan-bayashi/ParallelWaveGAN.
examples/ljspeech/voc5/conf/default.yaml 0 → 100644
浏览文件 @ c4035f8c
# This is the configuration file for LJSpeech dataset.
# This configuration is based on HiFiGAN V1, which is an official configuration.
# But I found that the optimizer setting does not work well with my implementation.
# So I changed optimizer settings as follows:
# - AdamW -> Adam
# - betas: [0.8, 0.99] -> betas: [0.5, 0.9]
# - Scheduler: ExponentialLR -> MultiStepLR
# To match the shift size difference, the upsample scales is also modified from the original 256 shift setting.
###########################################################
# FEATURE EXTRACTION SETTING #
###########################################################
fs: 22050 # Sampling rate.
n_fft: 1024 # FFT size (samples).
n_shift: 256 # Hop size (samples). 11.6ms
win_length: null # Window length (samples).
# If set to null, it will be the same as fft_size.
window: "hann" # Window function.
n_mels: 80 # Number of mel basis.
fmin: 80 # Minimum freq in mel basis calculation. (Hz)
fmax: 7600 # Maximum frequency in mel basis calculation. (Hz)
###########################################################
# GENERATOR NETWORK ARCHITECTURE SETTING #
###########################################################
generator_params:
in_channels: 80 # Number of input channels.
out_channels: 1 # Number of output channels.
channels: 512 # Number of initial channels.
kernel_size: 7 # Kernel size of initial and final conv layers.
upsample_scales: [8, 8, 2, 2] # Upsampling scales.
upsample_kernel_sizes: [16, 16, 4, 4] # Kernel size for upsampling layers.
resblock_kernel_sizes: [3, 7, 11] # Kernel size for residual blocks.
resblock_dilations: # Dilations for residual blocks.
- [1, 3, 5]
- [1, 3, 5]
- [1, 3, 5]
use_additional_convs: True # Whether to use additional conv layer in residual blocks.
bias: True # Whether to use bias parameter in conv.
nonlinear_activation: "leakyrelu" # Nonlinear activation type.
nonlinear_activation_params: # Nonlinear activation paramters.
negative_slope: 0.1
use_weight_norm: True # Whether to apply weight normalization.
###########################################################
# DISCRIMINATOR NETWORK ARCHITECTURE SETTING #
###########################################################
discriminator_params:
scales: 3 # Number of multi-scale discriminator.
scale_downsample_pooling: "AvgPool1D" # Pooling operation for scale discriminator.
scale_downsample_pooling_params:
kernel_size: 4 # Pooling kernel size.
stride: 2 # Pooling stride.
padding: 2 # Padding size.
scale_discriminator_params:
in_channels: 1 # Number of input channels.
out_channels: 1 # Number of output channels.
kernel_sizes: [15, 41, 5, 3] # List of kernel sizes.
channels: 128 # Initial number of channels.
max_downsample_channels: 1024 # Maximum number of channels in downsampling conv layers.
max_groups: 16 # Maximum number of groups in downsampling conv layers.
bias: True
downsample_scales: [4, 4, 4, 4, 1] # Downsampling scales.
nonlinear_activation: "leakyrelu" # Nonlinear activation.
nonlinear_activation_params:
negative_slope: 0.1
follow_official_norm: True # Whether to follow the official norm setting.
periods: [2, 3, 5, 7, 11] # List of period for multi-period discriminator.
period_discriminator_params:
in_channels: 1 # Number of input channels.
out_channels: 1 # Number of output channels.
kernel_sizes: [5, 3] # List of kernel sizes.
channels: 32 # Initial number of channels.
downsample_scales: [3, 3, 3, 3, 1] # Downsampling scales.
max_downsample_channels: 1024 # Maximum number of channels in downsampling conv layers.
bias: True # Whether to use bias parameter in conv layer."
nonlinear_activation: "leakyrelu" # Nonlinear activation.
nonlinear_activation_params: # Nonlinear activation paramters.
negative_slope: 0.1
use_weight_norm: True # Whether to apply weight normalization.
use_spectral_norm: False # Whether to apply spectral normalization.
###########################################################
# STFT LOSS SETTING #
###########################################################
use_stft_loss: False # Whether to use multi-resolution STFT loss.
use_mel_loss: True # Whether to use Mel-spectrogram loss.
mel_loss_params:
fs: 22050
fft_size: 1024
hop_size: 256
win_length: null
window: "hann"
num_mels: 80
fmin: 0
fmax: 11025
log_base: null
generator_adv_loss_params:
average_by_discriminators: False # Whether to average loss by #discriminators.
discriminator_adv_loss_params:
average_by_discriminators: False # Whether to average loss by #discriminators.
use_feat_match_loss: True
feat_match_loss_params:
average_by_discriminators: False # Whether to average loss by #discriminators.
average_by_layers: False # Whether to average loss by #layers in each discriminator.
include_final_outputs: False # Whether to include final outputs in feat match loss calculation.
###########################################################
# ADVERSARIAL LOSS SETTING #
###########################################################
lambda_aux: 45.0 # Loss balancing coefficient for STFT loss.
lambda_adv: 1.0 # Loss balancing coefficient for adversarial loss.
lambda_feat_match: 2.0 # Loss balancing coefficient for feat match loss..
###########################################################
# DATA LOADER SETTING #
###########################################################
batch_size: 16 # Batch size.
batch_max_steps: 8192 # Length of each audio in batch. Make sure dividable by hop_size.
num_workers: 2 # Number of workers in DataLoader.
###########################################################
# OPTIMIZER & SCHEDULER SETTING #
###########################################################
generator_optimizer_params:
beta1: 0.5
beta2: 0.9
weight_decay: 0.0 # Generator's weight decay coefficient.
generator_scheduler_params:
learning_rate: 2.0e-4 # Generator's learning rate.
gamma: 0.5 # Generator's scheduler gamma.
milestones: # At each milestone, lr will be multiplied by gamma.
- 200000
- 400000
- 600000
- 800000
generator_grad_norm: -1 # Generator's gradient norm.
discriminator_optimizer_params:
beta1: 0.5
beta2: 0.9
weight_decay: 0.0 # Discriminator's weight decay coefficient.
discriminator_scheduler_params:
learning_rate: 2.0e-4 # Discriminator's learning rate.
gamma: 0.5 # Discriminator's scheduler gamma.
milestones: # At each milestone, lr will be multiplied by gamma.
- 200000
- 400000
- 600000
- 800000
discriminator_grad_norm: -1 # Discriminator's gradient norm.
###########################################################
# INTERVAL SETTING #
###########################################################
generator_train_start_steps: 1 # Number of steps to start to train discriminator.
discriminator_train_start_steps: 0 # Number of steps to start to train discriminator.
train_max_steps: 2500000 # Number of training steps.
save_interval_steps: 5000 # Interval steps to save checkpoint.
eval_interval_steps: 1000 # Interval steps to evaluate the network.
###########################################################
# OTHER SETTING #
###########################################################
num_snapshots: 10 # max number of snapshots to keep while training
seed: 42 # random seed for paddle, random, and np.random
examples/ljspeech/voc5/local/preprocess.sh 0 → 100755
浏览文件 @ c4035f8c
#!/bin/bash
stage=0
stop_stage=100
config_path=$1
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
# get durations from MFA's result
echo "Generate durations.txt from MFA results ..."
python3 ${MAIN_ROOT}/utils/gen_duration_from_textgrid.py \
--inputdir=./ljspeech_alignment \
--output=durations.txt \
--config=${config_path}
fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# extract features
echo "Extract features ..."
python3 ${BIN_DIR}/../preprocess.py \
--rootdir=~/datasets/LJSpeech-1.1/ \
--dataset=ljspeech \
--dumpdir=dump \
--dur-file=durations.txt \
--config=${config_path} \
--cut-sil=True \
--num-cpu=20
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, 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 \
--stats=dump/train/feats_stats.npy
python3 ${BIN_DIR}/../normalize.py \
--metadata=dump/dev/raw/metadata.jsonl \
--dumpdir=dump/dev/norm \
--stats=dump/train/feats_stats.npy
python3 ${BIN_DIR}/../normalize.py \
--metadata=dump/test/raw/metadata.jsonl \
--dumpdir=dump/test/norm \
--stats=dump/train/feats_stats.npy
fi
examples/ljspeech/voc5/local/synthesize.sh 0 → 100755
浏览文件 @ c4035f8c
#!/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 ${BIN_DIR}/../synthesize.py \
--config=${config_path} \
--checkpoint=${train_output_path}/checkpoints/${ckpt_name} \
--test-metadata=dump/test/norm/metadata.jsonl \
--output-dir=${train_output_path}/test \
--generator-type=hifigan
examples/ljspeech/voc5/local/train.sh 0 → 100755
浏览文件 @ c4035f8c
#!/bin/bash
config_path=$1
train_output_path=$2
FLAGS_cudnn_exhaustive_search=true \
FLAGS_conv_workspace_size_limit=4000 \
python ${BIN_DIR}/train.py \
--train-metadata=dump/train/norm/metadata.jsonl \
--dev-metadata=dump/dev/norm/metadata.jsonl \
--config=${config_path} \
--output-dir=${train_output_path} \
--ngpu=1
examples/ljspeech/voc5/path.sh 0 → 100755
浏览文件 @ c4035f8c
#!/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=hifigan
export BIN_DIR=${MAIN_ROOT}/paddlespeech/t2s/exps/gan_vocoder/${MODEL}
examples/ljspeech/voc5/run.sh 0 → 100755
浏览文件 @ c4035f8c
#!/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_5000.pdz
# 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} || 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 ] && [ ${stop_stage} -ge 2 ]; then
# synthesize
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1
fi
examples/vctk/voc5/README.md 0 → 100644
浏览文件 @ c4035f8c
# HiFiGAN with VCTK
This example contains code used to train a [HiFiGAN](https://arxiv.org/abs/2010.05646) model with [VCTK](https://datashare.ed.ac.uk/handle/10283/3443).
## Dataset
### Download and Extract
Download VCTK-0.92 from the [official website](https://datashare.ed.ac.uk/handle/10283/3443) and extract it to `~/datasets`. Then the dataset is in directory `~/datasets/VCTK-Corpus-0.92`.
### Get MFA Result and Extract
We use [MFA](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) results to cut the silence in the edge of audio.
You can download from here [vctk_alignment.tar.gz](https://paddlespeech.bj.bcebos.com/MFA/VCTK-Corpus-0.92/vctk_alignment.tar.gz), or train your MFA model reference to [mfa example](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/mfa) of our repo.
ps: we remove three speakers in VCTK-0.92 (see [reorganize_vctk.py](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/examples/other/mfa/local/reorganize_vctk.py)):
1. `p315`, because of no text for it.
2. `p280` and `p362`, because no *_mic2.flac (which is better than *_mic1.flac) for them.
## Get Started
Assume the path to the dataset is `~/datasets/VCTK-Corpus-0.92`.
Assume the path to the MFA result of VCTK is `./vctk_alignment`.
Run the command below to
1. **source path**.
2. preprocess the dataset.
3. train the model.
4. synthesize wavs.
- synthesize waveform from `metadata.jsonl`.
```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 ${conf_path}
```
When it is done. A `dump` folder is created in the current directory. The structure of the dump folder is listed below.
```text
dump
├── dev
│ ├── norm
│ └── raw
├── test
│ ├── norm
│ └── raw
└── train
├── norm
├── raw
└── feats_stats.npy
```
The dataset is split into 3 parts, namely `train`, `dev`, and `test`, each of which contains a `norm` and `raw` subfolder. The `raw` folder contains the log magnitude of the mel spectrogram of each utterance, while the norm folder contains the normalized spectrogram. The statistics used to normalize the spectrogram are computed from the training set, which is located in `dump/train/feats_stats.npy`.
Also, there is a `metadata.jsonl` in each subfolder. It is a table-like file that contains id and paths to the spectrogram of each utterance.
### Model Training
```bash
CUDA_VISIBLE_DEVICES=${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] [--batch-size BATCH_SIZE] [--max-iter MAX_ITER]
[--run-benchmark RUN_BENCHMARK]
[--profiler_options PROFILER_OPTIONS]
Train a ParallelWaveGAN 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.
benchmark:
arguments related to benchmark.
--batch-size BATCH_SIZE
batch size.
--max-iter MAX_ITER train max steps.
--run-benchmark RUN_BENCHMARK
runing benchmark or not, if True, use the --batch-size
and --max-iter.
--profiler_options PROFILER_OPTIONS
The option of profiler, which should be in format
"key1=value1;key2=value2;key3=value3".
```
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.
### 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] [--generator-type GENERATOR_TYPE] [--config CONFIG]
[--checkpoint CHECKPOINT] [--test-metadata TEST_METADATA]
[--output-dir OUTPUT_DIR] [--ngpu NGPU]
Synthesize with GANVocoder.
optional arguments:
-h, --help show this help message and exit
--generator-type GENERATOR_TYPE
type of GANVocoder, should in {pwgan, mb_melgan,
style_melgan, } now
--config CONFIG GANVocoder config file.
--checkpoint CHECKPOINT
snapshot to load.
--test-metadata TEST_METADATA
dev data.
--output-dir OUTPUT_DIR
output dir.
--ngpu NGPU if ngpu == 0, use cpu.
```
1. `--config` config file. You should use the same config with which the model is trained.
2. `--checkpoint` is the checkpoint to load. Pick one of the checkpoints from `checkpoints` inside the training output directory.
3. `--test-metadata` is the metadata of the test dataset. Use the `metadata.jsonl` in the `dev/norm` subfolder from the processed directory.
4. `--output-dir` is the directory to save the synthesized audio files.
5. `--ngpu` is the number of gpus to use, if ngpu == 0, use cpu.
## Pretrained Model
## Acknowledgement
We adapted some code from https://github.com/kan-bayashi/ParallelWaveGAN.
examples/vctk/voc5/conf/default.yaml 0 → 100644
浏览文件 @ c4035f8c
# This is the configuration file for VCTK dataset.
# This configuration is based on HiFiGAN V1, which is
# an official configuration. But I found that the optimizer
# setting does not work well with my implementation.
# So I changed optimizer settings as follows:
# - AdamW -> Adam
# - betas: [0.8, 0.99] -> betas: [0.5, 0.9]
# - Scheduler: ExponentialLR -> MultiStepLR
# To match the shift size difference, the upsample scales
# is also modified from the original 256 shift setting.
###########################################################
# FEATURE EXTRACTION SETTING #
###########################################################
fs: 24000 # Sampling rate.
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.
n_mels: 80 # Number of mel basis.
fmin: 80 # Minimum freq in mel basis calculation. (Hz)
fmax: 7600 # Maximum frequency in mel basis calculation. (Hz)
###########################################################
# GENERATOR NETWORK ARCHITECTURE SETTING #
###########################################################
generator_params:
in_channels: 80 # Number of input channels.
out_channels: 1 # Number of output channels.
channels: 512 # Number of initial channels.
kernel_size: 7 # Kernel size of initial and final conv layers.
upsample_scales: [5, 5, 4, 3] # Upsampling scales.
upsample_kernel_sizes: [10, 10, 8, 6] # Kernel size for upsampling layers.
resblock_kernel_sizes: [3, 7, 11] # Kernel size for residual blocks.
resblock_dilations: # Dilations for residual blocks.
- [1, 3, 5]
- [1, 3, 5]
- [1, 3, 5]
use_additional_convs: True # Whether to use additional conv layer in residual blocks.
bias: True # Whether to use bias parameter in conv.
nonlinear_activation: "leakyrelu" # Nonlinear activation type.
nonlinear_activation_params: # Nonlinear activation paramters.
negative_slope: 0.1
use_weight_norm: True # Whether to apply weight normalization.
###########################################################
# DISCRIMINATOR NETWORK ARCHITECTURE SETTING #
###########################################################
discriminator_params:
scales: 3 # Number of multi-scale discriminator.
scale_downsample_pooling: "AvgPool1D" # Pooling operation for scale discriminator.
scale_downsample_pooling_params:
kernel_size: 4 # Pooling kernel size.
stride: 2 # Pooling stride.
padding: 2 # Padding size.
scale_discriminator_params:
in_channels: 1 # Number of input channels.
out_channels: 1 # Number of output channels.
kernel_sizes: [15, 41, 5, 3] # List of kernel sizes.
channels: 128 # Initial number of channels.
max_downsample_channels: 1024 # Maximum number of channels in downsampling conv layers.
max_groups: 16 # Maximum number of groups in downsampling conv layers.
bias: True
downsample_scales: [4, 4, 4, 4, 1] # Downsampling scales.
nonlinear_activation: "leakyrelu" # Nonlinear activation.
nonlinear_activation_params:
negative_slope: 0.1
follow_official_norm: True # Whether to follow the official norm setting.
periods: [2, 3, 5, 7, 11] # List of period for multi-period discriminator.
period_discriminator_params:
in_channels: 1 # Number of input channels.
out_channels: 1 # Number of output channels.
kernel_sizes: [5, 3] # List of kernel sizes.
channels: 32 # Initial number of channels.
downsample_scales: [3, 3, 3, 3, 1] # Downsampling scales.
max_downsample_channels: 1024 # Maximum number of channels in downsampling conv layers.
bias: True # Whether to use bias parameter in conv layer."
nonlinear_activation: "leakyrelu" # Nonlinear activation.
nonlinear_activation_params: # Nonlinear activation paramters.
negative_slope: 0.1
use_weight_norm: True # Whether to apply weight normalization.
use_spectral_norm: False # Whether to apply spectral normalization.
###########################################################
# STFT LOSS SETTING #
###########################################################
use_stft_loss: False # Whether to use multi-resolution STFT loss.
use_mel_loss: True # Whether to use Mel-spectrogram loss.
mel_loss_params:
fs: 24000
fft_size: 2048
hop_size: 300
win_length: 1200
window: "hann"
num_mels: 80
fmin: 0
fmax: 12000
log_base: null
generator_adv_loss_params:
average_by_discriminators: False # Whether to average loss by #discriminators.
discriminator_adv_loss_params:
average_by_discriminators: False # Whether to average loss by #discriminators.
use_feat_match_loss: True
feat_match_loss_params:
average_by_discriminators: False # Whether to average loss by #discriminators.
average_by_layers: False # Whether to average loss by #layers in each discriminator.
include_final_outputs: False # Whether to include final outputs in feat match loss calculation.
###########################################################
# ADVERSARIAL LOSS SETTING #
###########################################################
lambda_aux: 45.0 # Loss balancing coefficient for STFT loss.
lambda_adv: 1.0 # Loss balancing coefficient for adversarial loss.
lambda_feat_match: 2.0 # Loss balancing coefficient for feat match loss..
###########################################################
# DATA LOADER SETTING #
###########################################################
batch_size: 16 # Batch size.
batch_max_steps: 8400 # Length of each audio in batch. Make sure dividable by hop_size.
num_workers: 2 # Number of workers in DataLoader.
###########################################################
# OPTIMIZER & SCHEDULER SETTING #
###########################################################
generator_optimizer_params:
beta1: 0.5
beta2: 0.9
weight_decay: 0.0 # Generator's weight decay coefficient.
generator_scheduler_params:
learning_rate: 2.0e-4 # Generator's learning rate.
gamma: 0.5 # Generator's scheduler gamma.
milestones: # At each milestone, lr will be multiplied by gamma.
- 200000
- 400000
- 600000
- 800000
generator_grad_norm: -1 # Generator's gradient norm.
discriminator_optimizer_params:
beta1: 0.5
beta2: 0.9
weight_decay: 0.0 # Discriminator's weight decay coefficient.
discriminator_scheduler_params:
learning_rate: 2.0e-4 # Discriminator's learning rate.
gamma: 0.5 # Discriminator's scheduler gamma.
milestones: # At each milestone, lr will be multiplied by gamma.
- 200000
- 400000
- 600000
- 800000
discriminator_grad_norm: -1 # Discriminator's gradient norm.
###########################################################
# INTERVAL SETTING #
###########################################################
generator_train_start_steps: 1 # Number of steps to start to train discriminator.
discriminator_train_start_steps: 0 # Number of steps to start to train discriminator.
train_max_steps: 2500000 # Number of training steps.
save_interval_steps: 5000 # Interval steps to save checkpoint.
eval_interval_steps: 1000 # Interval steps to evaluate the network.
###########################################################
# OTHER SETTING #
###########################################################
num_snapshots: 10 # max number of snapshots to keep while training
seed: 42 # random seed for paddle, random, and np.random
examples/vctk/voc5/local/preprocess.sh 0 → 100755
浏览文件 @ c4035f8c
#!/bin/bash
stage=0
stop_stage=100
config_path=$1
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
# get durations from MFA's result
echo "Generate durations.txt from MFA results ..."
python3 ${MAIN_ROOT}/utils/gen_duration_from_textgrid.py \
--inputdir=./vctk_alignment \
--output=durations.txt \
--config=${config_path}
fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# extract features
echo "Extract features ..."
python3 ${BIN_DIR}/../preprocess.py \
--rootdir=~/datasets/VCTK-Corpus-0.92/ \
--dataset=vctk \
--dumpdir=dump \
--dur-file=durations.txt \
--config=${config_path} \
--cut-sil=True \
--num-cpu=20
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, 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 \
--stats=dump/train/feats_stats.npy
python3 ${BIN_DIR}/../normalize.py \
--metadata=dump/dev/raw/metadata.jsonl \
--dumpdir=dump/dev/norm \
--stats=dump/train/feats_stats.npy
python3 ${BIN_DIR}/../normalize.py \
--metadata=dump/test/raw/metadata.jsonl \
--dumpdir=dump/test/norm \
--stats=dump/train/feats_stats.npy
fi
examples/vctk/voc5/local/synthesize.sh 0 → 100755
浏览文件 @ c4035f8c
#!/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 ${BIN_DIR}/../synthesize.py \
--config=${config_path} \
--checkpoint=${train_output_path}/checkpoints/${ckpt_name} \
--test-metadata=dump/test/norm/metadata.jsonl \
--output-dir=${train_output_path}/test \
--generator-type=hifigan
examples/vctk/voc5/local/train.sh 0 → 100755
浏览文件 @ c4035f8c
#!/bin/bash
config_path=$1
train_output_path=$2
FLAGS_cudnn_exhaustive_search=true \
FLAGS_conv_workspace_size_limit=4000 \
python ${BIN_DIR}/train.py \
--train-metadata=dump/train/norm/metadata.jsonl \
--dev-metadata=dump/dev/norm/metadata.jsonl \
--config=${config_path} \
--output-dir=${train_output_path} \
--ngpu=1
examples/vctk/voc5/path.sh 0 → 100755
浏览文件 @ c4035f8c
#!/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=hifigan
export BIN_DIR=${MAIN_ROOT}/paddlespeech/t2s/exps/gan_vocoder/${MODEL}
examples/vctk/voc5/run.sh 0 → 100755
浏览文件 @ c4035f8c
#!/bin/bash
set -e
source path.sh
gpus=0
stage=0
stop_stage=100
conf_path=conf/default.yaml
train_output_path=exp/default
ckpt_name=snapshot_iter_5000.pdz
# 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} || 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 ] && [ ${stop_stage} -ge 2 ]; then
# synthesize
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1
fi
paddleaudio/CHANGELOG.md
浏览文件 @ c4035f8c
# Changelog
Date: 2022-3-15, Author: Xiaojie Chen.
- kaldi and librosa mfcc, fbank, spectrogram.
- unit test and benchmark.
Date: 2022-2-25, Author: Hui Zhang.
- Refactor architecture.
- dtw distance and mcd style dtw.
paddleaudio/features/augment.py 已删除 100644 → 0
浏览文件 @ 8d919781
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import List
import numpy as np
from numpy import ndarray as array
from ..backends import depth_convert
from ..utils import ParameterError
__all__ = [
'depth_augment',
'spect_augment',
'random_crop1d',
'random_crop2d',
'adaptive_spect_augment',
]
def randint(high: int) -> int:
"""Generate one random integer in range [0 high)
This is a helper function for random data augmentaiton
"""
return int(np.random.randint(0, high=high))
def rand() -> float:
"""Generate one floating-point number in range [0 1)
This is a helper function for random data augmentaiton
"""
return float(np.random.rand(1))
def depth_augment(y: array,
choices: List=['int8', 'int16'],
probs: List[float]=[0.5, 0.5]) -> array:
""" Audio depth augmentation
Do audio depth augmentation to simulate the distortion brought by quantization.
"""
assert len(probs) == len(
choices
), 'number of choices {} must be equal to size of probs {}'.format(
len(choices), len(probs))
depth = np.random.choice(choices, p=probs)
src_depth = y.dtype
y1 = depth_convert(y, depth)
y2 = depth_convert(y1, src_depth)
return y2
def adaptive_spect_augment(spect: array, tempo_axis: int=0,
level: float=0.1) -> array:
"""Do adpative spectrogram augmentation
The level of the augmentation is gowern by the paramter level,
ranging from 0 to 1, with 0 represents no augmentation。
"""
assert spect.ndim == 2., 'only supports 2d tensor or numpy array'
if tempo_axis == 0:
nt, nf = spect.shape
else:
nf, nt = spect.shape
time_mask_width = int(nt * level * 0.5)
freq_mask_width = int(nf * level * 0.5)
num_time_mask = int(10 * level)
num_freq_mask = int(10 * level)
if tempo_axis == 0:
for _ in range(num_time_mask):
start = randint(nt - time_mask_width)
spect[start:start + time_mask_width, :] = 0
for _ in range(num_freq_mask):
start = randint(nf - freq_mask_width)
spect[:, start:start + freq_mask_width] = 0
else:
for _ in range(num_time_mask):
start = randint(nt - time_mask_width)
spect[:, start:start + time_mask_width] = 0
for _ in range(num_freq_mask):
start = randint(nf - freq_mask_width)
spect[start:start + freq_mask_width, :] = 0
return spect
def spect_augment(spect: array,
tempo_axis: int=0,
max_time_mask: int=3,
max_freq_mask: int=3,
max_time_mask_width: int=30,
max_freq_mask_width: int=20) -> array:
"""Do spectrogram augmentation in both time and freq axis
Reference:
"""
assert spect.ndim == 2., 'only supports 2d tensor or numpy array'
if tempo_axis == 0:
nt, nf = spect.shape
else:
nf, nt = spect.shape
num_time_mask = randint(max_time_mask)
num_freq_mask = randint(max_freq_mask)
time_mask_width = randint(max_time_mask_width)
freq_mask_width = randint(max_freq_mask_width)
if tempo_axis == 0:
for _ in range(num_time_mask):
start = randint(nt - time_mask_width)
spect[start:start + time_mask_width, :] = 0
for _ in range(num_freq_mask):
start = randint(nf - freq_mask_width)
spect[:, start:start + freq_mask_width] = 0
else:
for _ in range(num_time_mask):
start = randint(nt - time_mask_width)
spect[:, start:start + time_mask_width] = 0
for _ in range(num_freq_mask):
start = randint(nf - freq_mask_width)
spect[start:start + freq_mask_width, :] = 0
return spect
def random_crop1d(y: array, crop_len: int) -> array:
""" Do random cropping on 1d input signal
The input is a 1d signal, typically a sound waveform
"""
if y.ndim != 1:
'only accept 1d tensor or numpy array'
n = len(y)
idx = randint(n - crop_len)
return y[idx:idx + crop_len]
def random_crop2d(s: array, crop_len: int, tempo_axis: int=0) -> array:
""" Do random cropping for 2D array, typically a spectrogram.
The cropping is done in temporal direction on the time-freq input signal.
"""
if tempo_axis >= s.ndim:
raise ParameterError('axis out of range')
n = s.shape[tempo_axis]
idx = randint(high=n - crop_len)
sli = [slice(None) for i in range(s.ndim)]
sli[tempo_axis] = slice(idx, idx + crop_len)
out = s[tuple(sli)]
return out
paddleaudio/paddleaudio/__init__.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from . import compliance
from . import datasets
from . import features
from . import functional
from . import io
from . import metric
from . import sox_effects
from .backends import load
from .backends import save
paddleaudio/paddleaudio/backends/__init__.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .soundfile_backend import depth_convert
from .soundfile_backend import load
from .soundfile_backend import normalize
from .soundfile_backend import resample
from .soundfile_backend import save
from .soundfile_backend import to_mono
paddleaudio/backends/audio.py → paddleaudio/paddleaudio/backends/soundfile_backend.py
浏览文件 @ c4035f8c
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -29,7 +29,7 @@ __all__ = [
'to_mono',
'depth_convert',
'normalize',
'save_wav',
'save',
'load',
]
NORMALMIZE_TYPES = ['linear', 'gaussian']
......@@ -41,12 +41,9 @@ EPS = 1e-8
def resample(y: array, src_sr: int, target_sr: int,
mode: str='kaiser_fast') -> array:
""" Audio resampling
This function is the same as using resampy.resample().
Notes:
The default mode is kaiser_fast. For better audio quality, use mode = 'kaiser_fast'
"""
if mode == 'kaiser_best':
......@@ -106,7 +103,6 @@ def to_mono(y: array, merge_type: str='average') -> array:
def _safe_cast(y: array, dtype: Union[type, str]) -> array:
""" data type casting in a safe way, i.e., prevent overflow or underflow
This function is used internally.
"""
return np.clip(y, np.iinfo(dtype).min, np.iinfo(dtype).max).astype(dtype)
......@@ -115,10 +111,8 @@ def _safe_cast(y: array, dtype: Union[type, str]) -> array:
def depth_convert(y: array, dtype: Union[type, str],
dithering: bool=True) -> array:
"""Convert audio array to target dtype safely
This function convert audio waveform to a target dtype, with addition steps of
preventing overflow/underflow and preserving audio range.
"""
SUPPORT_DTYPE = ['int16', 'int8', 'float32', 'float64']
......@@ -168,12 +162,9 @@ def sound_file_load(file: str,
dtype: str='int16',
duration: Optional[int]=None) -> Tuple[array, int]:
"""Load audio using soundfile library
This function load audio file using libsndfile.
Reference:
http://www.mega-nerd.com/libsndfile/#Features
"""
with sf.SoundFile(file) as sf_desc:
sr_native = sf_desc.samplerate
......@@ -188,33 +179,9 @@ def sound_file_load(file: str,
return y, sf_desc.samplerate
def audio_file_load():
"""Load audio using audiofile library
This function load audio file using audiofile.
Reference:
https://audiofile.68k.org/
"""
raise NotImplementedError()
def sox_file_load():
"""Load audio using sox library
This function load audio file using sox.
Reference:
http://sox.sourceforge.net/
"""
raise NotImplementedError()
def normalize(y: array, norm_type: str='linear',
mul_factor: float=1.0) -> array:
""" normalize an input audio with additional multiplier.
"""
if norm_type == 'linear':
......@@ -232,14 +199,12 @@ def normalize(y: array, norm_type: str='linear',
return y
def save_wav(y: array, sr: int, file: str) -> None:
def save(y: array, sr: int, file: str) -> None:
"""Save audio file to disk.
This function saves audio to disk using scipy.io.wavfile, with additional step
to convert input waveform to int16 unless it already is int16
Notes:
It only support raw wav format.
"""
if not file.endswith('.wav'):
raise ParameterError(
......@@ -274,11 +239,8 @@ def load(
resample_mode: str='kaiser_fast') -> Tuple[array, int]:
"""Load audio file from disk.
This function loads audio from disk using using audio beackend.
Parameters:
Notes:
"""
y, r = sound_file_load(file, offset=offset, dtype=dtype, duration=duration)
......
paddleaudio/paddleaudio/backends/sox_backend.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
paddleaudio/__init__.py → paddleaudio/paddleaudio/compliance/__init__.py
浏览文件 @ c4035f8c
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -11,5 +11,3 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .backends import *
from .features import *
paddleaudio/paddleaudio/compliance/kaldi.py 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
paddleaudio/features/core.py → paddleaudio/paddleaudio/compliance/librosa.py
浏览文件 @ c4035f8c
......@@ -21,11 +21,13 @@ import numpy as np
import scipy
from numpy import ndarray as array
from numpy.lib.stride_tricks import as_strided
from scipy.signal import get_window
from scipy import signal
from ..backends import depth_convert
from ..utils import ParameterError
__all__ = [
# dsp
'stft',
'mfcc',
'hz_to_mel',
......@@ -38,6 +40,12 @@ __all__ = [
'spectrogram',
'mu_encode',
'mu_decode',
# augmentation
'depth_augment',
'spect_augment',
'random_crop1d',
'random_crop2d',
'adaptive_spect_augment',
]
......@@ -303,7 +311,7 @@ def stft(x: array,
if hop_length is None:
hop_length = int(win_length // 4)
fft_window = get_window(window, win_length, fftbins=True)
fft_window = signal.get_window(window, win_length, fftbins=True)
# Pad the window out to n_fft size
fft_window = pad_center(fft_window, n_fft)
......@@ -576,3 +584,145 @@ def mu_decode(y: array, mu: int=255, quantized: bool=True) -> array:
y = y * 2 / mu - 1
x = np.sign(y) / mu * ((1 + mu)**np.abs(y) - 1)
return x
def randint(high: int) -> int:
"""Generate one random integer in range [0 high)
This is a helper function for random data augmentaiton
"""
return int(np.random.randint(0, high=high))
def rand() -> float:
"""Generate one floating-point number in range [0 1)
This is a helper function for random data augmentaiton
"""
return float(np.random.rand(1))
def depth_augment(y: array,
choices: List=['int8', 'int16'],
probs: List[float]=[0.5, 0.5]) -> array:
""" Audio depth augmentation
Do audio depth augmentation to simulate the distortion brought by quantization.
"""
assert len(probs) == len(
choices
), 'number of choices {} must be equal to size of probs {}'.format(
len(choices), len(probs))
depth = np.random.choice(choices, p=probs)
src_depth = y.dtype
y1 = depth_convert(y, depth)
y2 = depth_convert(y1, src_depth)
return y2
def adaptive_spect_augment(spect: array, tempo_axis: int=0,
level: float=0.1) -> array:
"""Do adpative spectrogram augmentation
The level of the augmentation is gowern by the paramter level,
ranging from 0 to 1, with 0 represents no augmentation。
"""
assert spect.ndim == 2., 'only supports 2d tensor or numpy array'
if tempo_axis == 0:
nt, nf = spect.shape
else:
nf, nt = spect.shape
time_mask_width = int(nt * level * 0.5)
freq_mask_width = int(nf * level * 0.5)
num_time_mask = int(10 * level)
num_freq_mask = int(10 * level)
if tempo_axis == 0:
for _ in range(num_time_mask):
start = randint(nt - time_mask_width)
spect[start:start + time_mask_width, :] = 0
for _ in range(num_freq_mask):
start = randint(nf - freq_mask_width)
spect[:, start:start + freq_mask_width] = 0
else:
for _ in range(num_time_mask):
start = randint(nt - time_mask_width)
spect[:, start:start + time_mask_width] = 0
for _ in range(num_freq_mask):
start = randint(nf - freq_mask_width)
spect[start:start + freq_mask_width, :] = 0
return spect
def spect_augment(spect: array,
tempo_axis: int=0,
max_time_mask: int=3,
max_freq_mask: int=3,
max_time_mask_width: int=30,
max_freq_mask_width: int=20) -> array:
"""Do spectrogram augmentation in both time and freq axis
Reference:
"""
assert spect.ndim == 2., 'only supports 2d tensor or numpy array'
if tempo_axis == 0:
nt, nf = spect.shape
else:
nf, nt = spect.shape
num_time_mask = randint(max_time_mask)
num_freq_mask = randint(max_freq_mask)
time_mask_width = randint(max_time_mask_width)
freq_mask_width = randint(max_freq_mask_width)
if tempo_axis == 0:
for _ in range(num_time_mask):
start = randint(nt - time_mask_width)
spect[start:start + time_mask_width, :] = 0
for _ in range(num_freq_mask):
start = randint(nf - freq_mask_width)
spect[:, start:start + freq_mask_width] = 0
else:
for _ in range(num_time_mask):
start = randint(nt - time_mask_width)
spect[:, start:start + time_mask_width] = 0
for _ in range(num_freq_mask):
start = randint(nf - freq_mask_width)
spect[start:start + freq_mask_width, :] = 0
return spect
def random_crop1d(y: array, crop_len: int) -> array:
""" Do random cropping on 1d input signal
The input is a 1d signal, typically a sound waveform
"""
if y.ndim != 1:
'only accept 1d tensor or numpy array'
n = len(y)
idx = randint(n - crop_len)
return y[idx:idx + crop_len]
def random_crop2d(s: array, crop_len: int, tempo_axis: int=0) -> array:
""" Do random cropping for 2D array, typically a spectrogram.
The cropping is done in temporal direction on the time-freq input signal.
"""
if tempo_axis >= s.ndim:
raise ParameterError('axis out of range')
n = s.shape[tempo_axis]
idx = randint(high=n - crop_len)
sli = [slice(None) for i in range(s.ndim)]
sli[tempo_axis] = slice(idx, idx + crop_len)
out = s[tuple(sli)]
return out
paddleaudio/datasets/__init__.py → paddleaudio/paddleaudio/datasets/__init__.py
浏览文件 @ c4035f8c
......@@ -15,10 +15,3 @@ from .esc50 import ESC50
from .gtzan import GTZAN
from .tess import TESS
from .urban_sound import UrbanSound8K
__all__ = [
'ESC50',
'UrbanSound8K',
'GTZAN',
'TESS',
]
paddleaudio/datasets/dataset.py → paddleaudio/paddleaudio/datasets/dataset.py
浏览文件 @ c4035f8c
......@@ -17,8 +17,8 @@ import numpy as np
import paddle
from ..backends import load as load_audio
from ..features import melspectrogram
from ..features import mfcc
from ..compliance.librosa import melspectrogram
from ..compliance.librosa import mfcc
feat_funcs = {
'raw': None,
......
paddleaudio/features/__init__.py → paddleaudio/paddleaudio/features/__init__.py
浏览文件 @ c4035f8c
......@@ -11,6 +11,7 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .augment import *
from .core import *
from .spectrum import *
from .layers import LogMelSpectrogram
from .layers import MelSpectrogram
from .layers import MFCC
from .layers import Spectrogram
paddleaudio/paddleaudio/functional/__init__.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .functional import compute_fbank_matrix
from .functional import create_dct
from .functional import fft_frequencies
from .functional import hz_to_mel
from .functional import mel_frequencies
from .functional import mel_to_hz
from .functional import power_to_db
paddleaudio/paddleaudio/functional/functional.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Modified from librosa(https://github.com/librosa/librosa)
import math
from typing import Optional
from typing import Union
import paddle
__all__ = [
'hz_to_mel',
'mel_to_hz',
'mel_frequencies',
'fft_frequencies',
'compute_fbank_matrix',
'power_to_db',
'create_dct',
]
def hz_to_mel(freq: Union[paddle.Tensor, float],
htk: bool=False) -> Union[paddle.Tensor, float]:
"""Convert Hz to Mels.
Parameters:
freq: the input tensor of arbitrary shape, or a single floating point number.
htk: use HTK formula to do the conversion.
The default value is False.
Returns:
The frequencies represented in Mel-scale.
"""
if htk:
if isinstance(freq, paddle.Tensor):
return 2595.0 * paddle.log10(1.0 + freq / 700.0)
else:
return 2595.0 * math.log10(1.0 + freq / 700.0)
# Fill in the linear part
f_min = 0.0
f_sp = 200.0 / 3
mels = (freq - f_min) / f_sp
# Fill in the log-scale part
min_log_hz = 1000.0 # beginning of log region (Hz)
min_log_mel = (min_log_hz - f_min) / f_sp # same (Mels)
logstep = math.log(6.4) / 27.0 # step size for log region
if isinstance(freq, paddle.Tensor):
target = min_log_mel + paddle.log(
freq / min_log_hz + 1e-10) / logstep # prevent nan with 1e-10
mask = (freq > min_log_hz).astype(freq.dtype)
mels = target * mask + mels * (
1 - mask) # will replace by masked_fill OP in future
else:
if freq >= min_log_hz:
mels = min_log_mel + math.log(freq / min_log_hz + 1e-10) / logstep
return mels
def mel_to_hz(mel: Union[float, paddle.Tensor],
htk: bool=False) -> Union[float, paddle.Tensor]:
"""Convert mel bin numbers to frequencies.
Parameters:
mel: the mel frequency represented as a tensor of arbitrary shape, or a floating point number.
htk: use HTK formula to do the conversion.
Returns:
The frequencies represented in hz.
"""
if htk:
return 700.0 * (10.0**(mel / 2595.0) - 1.0)
f_min = 0.0
f_sp = 200.0 / 3
freqs = f_min + f_sp * mel
# And now the nonlinear scale
min_log_hz = 1000.0 # beginning of log region (Hz)
min_log_mel = (min_log_hz - f_min) / f_sp # same (Mels)
logstep = math.log(6.4) / 27.0 # step size for log region
if isinstance(mel, paddle.Tensor):
target = min_log_hz * paddle.exp(logstep * (mel - min_log_mel))
mask = (mel > min_log_mel).astype(mel.dtype)
freqs = target * mask + freqs * (
1 - mask) # will replace by masked_fill OP in future
else:
if mel >= min_log_mel:
freqs = min_log_hz * math.exp(logstep * (mel - min_log_mel))
return freqs
def mel_frequencies(n_mels: int=64,
f_min: float=0.0,
f_max: float=11025.0,
htk: bool=False,
dtype: str=paddle.float32):
"""Compute mel frequencies.
Parameters:
n_mels(int): number of Mel bins.
f_min(float): the lower cut-off frequency, below which the filter response is zero.
f_max(float): the upper cut-off frequency, above which the filter response is zero.
htk(bool): whether to use htk formula.
dtype(str): the datatype of the return frequencies.
Returns:
The frequencies represented in Mel-scale
"""
# 'Center freqs' of mel bands - uniformly spaced between limits
min_mel = hz_to_mel(f_min, htk=htk)
max_mel = hz_to_mel(f_max, htk=htk)
mels = paddle.linspace(min_mel, max_mel, n_mels, dtype=dtype)
freqs = mel_to_hz(mels, htk=htk)
return freqs
def fft_frequencies(sr: int, n_fft: int, dtype: str=paddle.float32):
"""Compute fourier frequencies.
Parameters:
sr(int): the audio sample rate.
n_fft(float): the number of fft bins.
dtype(str): the datatype of the return frequencies.
Returns:
The frequencies represented in hz.
"""
return paddle.linspace(0, float(sr) / 2, int(1 + n_fft // 2), dtype=dtype)
def compute_fbank_matrix(sr: int,
n_fft: int,
n_mels: int=64,
f_min: float=0.0,
f_max: Optional[float]=None,
htk: bool=False,
norm: Union[str, float]='slaney',
dtype: str=paddle.float32):
"""Compute fbank matrix.
Parameters:
sr(int): the audio sample rate.
n_fft(int): the number of fft bins.
n_mels(int): the number of Mel bins.
f_min(float): the lower cut-off frequency, below which the filter response is zero.
f_max(float): the upper cut-off frequency, above which the filter response is zero.
htk: whether to use htk formula.
return_complex(bool): whether to return complex matrix. If True, the matrix will
be complex type. Otherwise, the real and image part will be stored in the last
axis of returned tensor.
dtype(str): the datatype of the returned fbank matrix.
Returns:
The fbank matrix of shape (n_mels, int(1+n_fft//2)).
Shape:
output: (n_mels, int(1+n_fft//2))
"""
if f_max is None:
f_max = float(sr) / 2
# Initialize the weights
weights = paddle.zeros((n_mels, int(1 + n_fft // 2)), dtype=dtype)
# Center freqs of each FFT bin
fftfreqs = fft_frequencies(sr=sr, n_fft=n_fft, dtype=dtype)
# 'Center freqs' of mel bands - uniformly spaced between limits
mel_f = mel_frequencies(
n_mels + 2, f_min=f_min, f_max=f_max, htk=htk, dtype=dtype)
fdiff = mel_f[1:] - mel_f[:-1] #np.diff(mel_f)
ramps = mel_f.unsqueeze(1) - fftfreqs.unsqueeze(0)
#ramps = np.subtract.outer(mel_f, fftfreqs)
for i in range(n_mels):
# lower and upper slopes for all bins
lower = -ramps[i] / fdiff[i]
upper = ramps[i + 2] / fdiff[i + 1]
# .. then intersect them with each other and zero
weights[i] = paddle.maximum(
paddle.zeros_like(lower), paddle.minimum(lower, upper))
# Slaney-style mel is scaled to be approx constant energy per channel
if norm == 'slaney':
enorm = 2.0 / (mel_f[2:n_mels + 2] - mel_f[:n_mels])
weights *= enorm.unsqueeze(1)
elif isinstance(norm, int) or isinstance(norm, float):
weights = paddle.nn.functional.normalize(weights, p=norm, axis=-1)
return weights
def power_to_db(magnitude: paddle.Tensor,
ref_value: float=1.0,
amin: float=1e-10,
top_db: Optional[float]=None) -> paddle.Tensor:
"""Convert a power spectrogram (amplitude squared) to decibel (dB) units.
The function computes the scaling ``10 * log10(x / ref)`` in a numerically
stable way.
Parameters:
magnitude(Tensor): the input magnitude tensor of any shape.
ref_value(float): the reference value. If smaller than 1.0, the db level
of the signal will be pulled up accordingly. Otherwise, the db level
is pushed down.
amin(float): the minimum value of input magnitude, below which the input
magnitude is clipped(to amin).
top_db(float): the maximum db value of resulting spectrum, above which the
spectrum is clipped(to top_db).
Returns:
The spectrogram in log-scale.
shape:
input: any shape
output: same as input
"""
if amin <= 0:
raise Exception("amin must be strictly positive")
if ref_value <= 0:
raise Exception("ref_value must be strictly positive")
ones = paddle.ones_like(magnitude)
log_spec = 10.0 * paddle.log10(paddle.maximum(ones * amin, magnitude))
log_spec -= 10.0 * math.log10(max(ref_value, amin))
if top_db is not None:
if top_db < 0:
raise Exception("top_db must be non-negative")
log_spec = paddle.maximum(log_spec, ones * (log_spec.max() - top_db))
return log_spec
def create_dct(n_mfcc: int,
n_mels: int,
norm: Optional[str]='ortho',
dtype: Optional[str]=paddle.float32) -> paddle.Tensor:
"""Create a discrete cosine transform(DCT) matrix.
Parameters:
n_mfcc (int): Number of mel frequency cepstral coefficients.
n_mels (int): Number of mel filterbanks.
norm (str, optional): Normalizaiton type. Defaults to 'ortho'.
Returns:
Tensor: The DCT matrix with shape (n_mels, n_mfcc).
"""
n = paddle.arange(n_mels, dtype=dtype)
k = paddle.arange(n_mfcc, dtype=dtype).unsqueeze(1)
dct = paddle.cos(math.pi / float(n_mels) * (n + 0.5) *
k) # size (n_mfcc, n_mels)
if norm is None:
dct *= 2.0
else:
assert norm == "ortho"
dct[0] *= 1.0 / math.sqrt(2.0)
dct *= math.sqrt(2.0 / float(n_mels))
return dct.T
paddleaudio/features/window.py → paddleaudio/paddleaudio/functional/window.py
浏览文件 @ c4035f8c
......@@ -20,6 +20,19 @@ from paddle import Tensor
__all__ = [
'get_window',
# windows
'taylor',
'hamming',
'hann',
'tukey',
'kaiser',
'gaussian',
'exponential',
'triang',
'bohman',
'blackman',
'cosine',
]
......@@ -73,6 +86,21 @@ def general_gaussian(M: int, p, sig, sym: bool=True,
return _truncate(w, needs_trunc)
def general_cosine(M: int, a: float, sym: bool=True,
dtype: str='float64') -> Tensor:
"""Compute a generic weighted sum of cosine terms window.
This function is consistent with scipy.signal.windows.general_cosine().
"""
if _len_guards(M):
return paddle.ones((M, ), dtype=dtype)
M, needs_trunc = _extend(M, sym)
fac = paddle.linspace(-math.pi, math.pi, M, dtype=dtype)
w = paddle.zeros((M, ), dtype=dtype)
for k in range(len(a)):
w += a[k] * paddle.cos(k * fac)
return _truncate(w, needs_trunc)
def general_hamming(M: int, alpha: float, sym: bool=True,
dtype: str='float64') -> Tensor:
"""Compute a generalized Hamming window.
......@@ -143,21 +171,6 @@ def taylor(M: int,
return _truncate(w, needs_trunc)
def general_cosine(M: int, a: float, sym: bool=True,
dtype: str='float64') -> Tensor:
"""Compute a generic weighted sum of cosine terms window.
This function is consistent with scipy.signal.windows.general_cosine().
"""
if _len_guards(M):
return paddle.ones((M, ), dtype=dtype)
M, needs_trunc = _extend(M, sym)
fac = paddle.linspace(-math.pi, math.pi, M, dtype=dtype)
w = paddle.zeros((M, ), dtype=dtype)
for k in range(len(a)):
w += a[k] * paddle.cos(k * fac)
return _truncate(w, needs_trunc)
def hamming(M: int, sym: bool=True, dtype: str='float64') -> Tensor:
"""Compute a Hamming window.
The Hamming window is a taper formed by using a raised cosine with
......@@ -375,6 +388,7 @@ def cosine(M: int, sym: bool=True, dtype: str='float64') -> Tensor:
return _truncate(w, needs_trunc)
## factory function
def get_window(window: Union[str, Tuple[str, float]],
win_length: int,
fftbins: bool=True,
......
paddleaudio/backends/__init__.py → paddleaudio/paddleaudio/io/__init__.py
浏览文件 @ c4035f8c
......@@ -11,4 +11,3 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .audio import *
paddleaudio/paddleaudio/metric/__init__.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .dtw import dtw_distance
from .mcd import mcd_distance
paddleaudio/paddleaudio/metric/dtw.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
from dtaidistance import dtw_ndim
__all__ = [
'dtw_distance',
]
def dtw_distance(xs: np.ndarray, ys: np.ndarray) -> float:
"""dtw distance
Dynamic Time Warping.
This function keeps a compact matrix, not the full warping paths matrix.
Uses dynamic programming to compute:
wps[i, j] = (s1[i]-s2[j])**2 + min(
wps[i-1, j ] + penalty, // vertical / insertion / expansion
wps[i , j-1] + penalty, // horizontal / deletion / compression
wps[i-1, j-1]) // diagonal / match
dtw = sqrt(wps[-1, -1])
Args:
xs (np.ndarray): ref sequence, [T,D]
ys (np.ndarray): hyp sequence, [T,D]
Returns:
float: dtw distance
"""
return dtw_ndim.distance(xs, ys)
paddleaudio/paddleaudio/metric/mcd.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import mcd.metrics_fast as mt
import numpy as np
from mcd import dtw
__all__ = [
'mcd_distance',
]
def mcd_distance(xs: np.ndarray, ys: np.ndarray, cost_fn=mt.logSpecDbDist):
"""Mel cepstral distortion (MCD), dtw distance.
Dynamic Time Warping.
Uses dynamic programming to compute:
wps[i, j] = cost_fn(xs[i], ys[j]) + min(
wps[i-1, j ], // vertical / insertion / expansion
wps[i , j-1], // horizontal / deletion / compression
wps[i-1, j-1]) // diagonal / match
dtw = sqrt(wps[-1, -1])
Cost Function:
logSpecDbConst = 10.0 / math.log(10.0) * math.sqrt(2.0)
def logSpecDbDist(x, y):
diff = x - y
return logSpecDbConst * math.sqrt(np.inner(diff, diff))
Args:
xs (np.ndarray): ref sequence, [T,D]
ys (np.ndarray): hyp sequence, [T,D]
Returns:
float: dtw distance
"""
min_cost, path = dtw.dtw(xs, ys, cost_fn)
return min_cost
paddleaudio/paddleaudio/sox_effects/__init__.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
paddleaudio/utils/__init__.py → paddleaudio/paddleaudio/utils/__init__.py
浏览文件 @ c4035f8c
......@@ -11,8 +11,15 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .download import *
from .env import *
from .error import *
from .log import *
from .time import *
from .download import decompress
from .download import download_and_decompress
from .download import load_state_dict_from_url
from .env import DATA_HOME
from .env import MODEL_HOME
from .env import PPAUDIO_HOME
from .env import USER_HOME
from .error import ParameterError
from .log import Logger
from .log import logger
from .time import seconds_to_hms
from .time import Timer
paddleaudio/utils/download.py → paddleaudio/paddleaudio/utils/download.py
浏览文件 @ c4035f8c
......@@ -22,6 +22,12 @@ from .log import logger
download.logger = logger
__all__ = [
'decompress',
'download_and_decompress',
'load_state_dict_from_url',
]
def decompress(file: str):
"""
......
paddleaudio/utils/env.py → paddleaudio/paddleaudio/utils/env.py
浏览文件 @ c4035f8c
......@@ -20,6 +20,13 @@ PPAUDIO_HOME --> the root directory for storing PaddleAudio related data. D
'''
import os
__all__ = [
'USER_HOME',
'PPAUDIO_HOME',
'MODEL_HOME',
'DATA_HOME',
]
def _get_user_home():
return os.path.expanduser('~')
......
paddleaudio/utils/log.py → paddleaudio/paddleaudio/utils/log.py
浏览文件 @ c4035f8c
......@@ -19,7 +19,10 @@ import time
import colorlog
loggers = {}
__all__ = [
'Logger',
'logger',
]
log_config = {
'DEBUG': {
......
paddleaudio/utils/time.py → paddleaudio/paddleaudio/utils/time.py
浏览文件 @ c4035f8c
......@@ -14,6 +14,11 @@
import math
import time
__all__ = [
'Timer',
'seconds_to_hms',
]
class Timer(object):
'''Calculate runing speed and estimated time of arrival(ETA)'''
......
setup_audio.py → paddleaudio/setup.py
浏览文件 @ c4035f8c
......@@ -11,19 +11,46 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import glob
import os
import setuptools
from setuptools.command.install import install
from setuptools.command.test import test
# set the version here
VERSION = '0.1.0'
VERSION = '0.2.0'
# Inspired by the example at https://pytest.org/latest/goodpractises.html
class TestCommand(test):
def finalize_options(self):
test.finalize_options(self)
self.test_args = []
self.test_suite = True
def run(self):
self.run_benchmark()
super(TestCommand, self).run()
def run_tests(self):
# Run nose ensuring that argv simulates running nosetests directly
import nose
nose.run_exit(argv=['nosetests', '-w', 'tests'])
def run_benchmark(self):
for benchmark_item in glob.glob('tests/benchmark/*py'):
os.system(f'pytest {benchmark_item}')
class InstallCommand(install):
def run(self):
install.run(self)
def write_version_py(filename='paddleaudio/__init__.py'):
import paddleaudio
if hasattr(paddleaudio,
"__version__") and paddleaudio.__version__ == VERSION:
return
with open(filename, "a") as f:
f.write(f"\n__version__ = '{VERSION}'\n")
f.write(f"__version__ = '{VERSION}'")
def remove_version_py(filename='paddleaudio/__init__.py'):
......@@ -35,6 +62,7 @@ def remove_version_py(filename='paddleaudio/__init__.py'):
f.write(line)
remove_version_py()
write_version_py()
setuptools.setup(
......@@ -59,6 +87,18 @@ setuptools.setup(
'resampy >= 0.2.2',
'soundfile >= 0.9.0',
'colorlog',
], )
'dtaidistance >= 2.3.6',
'mcd >= 0.4',
],
extras_require={
'test': [
'nose', 'librosa==0.8.1', 'soundfile==0.10.3.post1',
'torchaudio==0.10.2', 'pytest-benchmark'
],
},
cmdclass={
'install': InstallCommand,
'test': TestCommand,
}, )
remove_version_py()
paddleaudio/tests/backends/__init__.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
paddleaudio/tests/backends/base.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import unittest
import urllib.request
mono_channel_wav = 'https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav'
multi_channels_wav = 'https://paddlespeech.bj.bcebos.com/PaddleAudio/cat.wav'
class BackendTest(unittest.TestCase):
def setUp(self):
self.initWavInput()
def initWavInput(self):
self.files = []
for url in [mono_channel_wav, multi_channels_wav]:
if not os.path.isfile(os.path.basename(url)):
urllib.request.urlretrieve(url, os.path.basename(url))
self.files.append(os.path.basename(url))
def initParmas(self):
raise NotImplementedError
paddleaudio/tests/backends/soundfile/__init__.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
paddleaudio/tests/backends/soundfile/test_io.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import filecmp
import os
import unittest
import numpy as np
import soundfile as sf
import paddleaudio
from ..base import BackendTest
class TestIO(BackendTest):
def test_load_mono_channel(self):
sf_data, sf_sr = sf.read(self.files[0])
pa_data, pa_sr = paddleaudio.load(
self.files[0], normal=False, dtype='float64')
self.assertEqual(sf_data.dtype, pa_data.dtype)
self.assertEqual(sf_sr, pa_sr)
np.testing.assert_array_almost_equal(sf_data, pa_data)
def test_load_multi_channels(self):
sf_data, sf_sr = sf.read(self.files[1])
sf_data = sf_data.T # Channel dim first
pa_data, pa_sr = paddleaudio.load(
self.files[1], mono=False, normal=False, dtype='float64')
self.assertEqual(sf_data.dtype, pa_data.dtype)
self.assertEqual(sf_sr, pa_sr)
np.testing.assert_array_almost_equal(sf_data, pa_data)
def test_save_mono_channel(self):
waveform, sr = np.random.randint(
low=-32768, high=32768, size=(48000), dtype=np.int16), 16000
sf_tmp_file = 'sf_tmp.wav'
pa_tmp_file = 'pa_tmp.wav'
sf.write(sf_tmp_file, waveform, sr)
paddleaudio.save(waveform, sr, pa_tmp_file)
self.assertTrue(filecmp.cmp(sf_tmp_file, pa_tmp_file))
for file in [sf_tmp_file, pa_tmp_file]:
os.remove(file)
def test_save_multi_channels(self):
waveform, sr = np.random.randint(
low=-32768, high=32768, size=(2, 48000), dtype=np.int16), 16000
sf_tmp_file = 'sf_tmp.wav'
pa_tmp_file = 'pa_tmp.wav'
sf.write(sf_tmp_file, waveform.T, sr)
paddleaudio.save(waveform.T, sr, pa_tmp_file)
self.assertTrue(filecmp.cmp(sf_tmp_file, pa_tmp_file))
for file in [sf_tmp_file, pa_tmp_file]:
os.remove(file)
if __name__ == '__main__':
unittest.main()
paddleaudio/tests/benchmark/README.md 0 → 100644
浏览文件 @ c4035f8c
# 1. Prepare
First, install `pytest-benchmark` via pip.
```sh
pip install pytest-benchmark
```
# 2. Run
Run the specific script for profiling.
```sh
pytest melspectrogram.py
```
Result:
```sh
========================================================================== test session starts ==========================================================================
platform linux -- Python 3.7.7, pytest-7.0.1, pluggy-1.0.0
benchmark: 3.4.1 (defaults: timer=time.perf_counter disable_gc=False min_rounds=5 min_time=0.000005 max_time=1.0 calibration_precision=10 warmup=False warmup_iterations=100000)
rootdir: /ssd3/chenxiaojie06/PaddleSpeech/DeepSpeech/paddleaudio
plugins: typeguard-2.12.1, benchmark-3.4.1, anyio-3.5.0
collected 4 items
melspectrogram.py .... [100%]
-------------------------------------------------------------------------------------------------- benchmark: 4 tests -------------------------------------------------------------------------------------------------
Name (time in us) Min Max Mean StdDev Median IQR Outliers OPS Rounds Iterations
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
test_melspect_gpu_torchaudio 202.0765 (1.0) 360.6230 (1.0) 218.1168 (1.0) 16.3022 (1.0) 214.2871 (1.0) 21.8451 (1.0) 40;3 4,584.7001 (1.0) 286 1
test_melspect_gpu 657.8509 (3.26) 908.0470 (2.52) 724.2545 (3.32) 106.5771 (6.54) 669.9096 (3.13) 113.4719 (5.19) 1;0 1,380.7300 (0.30) 5 1
test_melspect_cpu_torchaudio 1,247.6053 (6.17) 2,892.5799 (8.02) 1,443.2853 (6.62) 345.3732 (21.19) 1,262.7263 (5.89) 221.6385 (10.15) 56;53 692.8637 (0.15) 399 1
test_melspect_cpu 20,326.2549 (100.59) 20,607.8682 (57.15) 20,473.4125 (93.86) 63.8654 (3.92) 20,467.0429 (95.51) 68.4294 (3.13) 8;1 48.8438 (0.01) 29 1
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Legend:
Outliers: 1 Standard Deviation from Mean; 1.5 IQR (InterQuartile Range) from 1st Quartile and 3rd Quartile.
OPS: Operations Per Second, computed as 1 / Mean
========================================================================== 4 passed in 21.12s ===========================================================================
```
paddleaudio/tests/benchmark/log_melspectrogram.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import urllib.request
import librosa
import numpy as np
import paddle
import torch
import torchaudio
import paddleaudio
wav_url = 'https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav'
if not os.path.isfile(os.path.basename(wav_url)):
urllib.request.urlretrieve(wav_url, os.path.basename(wav_url))
waveform, sr = paddleaudio.load(os.path.abspath(os.path.basename(wav_url)))
waveform_tensor = paddle.to_tensor(waveform).unsqueeze(0)
waveform_tensor_torch = torch.from_numpy(waveform).unsqueeze(0)
# Feature conf
mel_conf = {
'sr': sr,
'n_fft': 512,
'hop_length': 128,
'n_mels': 40,
}
mel_conf_torchaudio = {
'sample_rate': sr,
'n_fft': 512,
'hop_length': 128,
'n_mels': 40,
'norm': 'slaney',
'mel_scale': 'slaney',
}
def enable_cpu_device():
paddle.set_device('cpu')
def enable_gpu_device():
paddle.set_device('gpu')
log_mel_extractor = paddleaudio.features.LogMelSpectrogram(
**mel_conf, f_min=0.0, top_db=80.0, dtype=waveform_tensor.dtype)
def log_melspectrogram():
return log_mel_extractor(waveform_tensor).squeeze(0)
def test_log_melspect_cpu(benchmark):
enable_cpu_device()
feature_paddleaudio = benchmark(log_melspectrogram)
feature_librosa = librosa.feature.melspectrogram(waveform, **mel_conf)
feature_librosa = librosa.power_to_db(feature_librosa, top_db=80.0)
np.testing.assert_array_almost_equal(
feature_librosa, feature_paddleaudio, decimal=3)
def test_log_melspect_gpu(benchmark):
enable_gpu_device()
feature_paddleaudio = benchmark(log_melspectrogram)
feature_librosa = librosa.feature.melspectrogram(waveform, **mel_conf)
feature_librosa = librosa.power_to_db(feature_librosa, top_db=80.0)
np.testing.assert_array_almost_equal(
feature_librosa, feature_paddleaudio, decimal=2)
mel_extractor_torchaudio = torchaudio.transforms.MelSpectrogram(
**mel_conf_torchaudio, f_min=0.0)
amplitude_to_DB = torchaudio.transforms.AmplitudeToDB('power', top_db=80.0)
def melspectrogram_torchaudio():
return mel_extractor_torchaudio(waveform_tensor_torch).squeeze(0)
def log_melspectrogram_torchaudio():
mel_specgram = mel_extractor_torchaudio(waveform_tensor_torch)
return amplitude_to_DB(mel_specgram).squeeze(0)
def test_log_melspect_cpu_torchaudio(benchmark):
global waveform_tensor_torch, mel_extractor_torchaudio, amplitude_to_DB
mel_extractor_torchaudio = mel_extractor_torchaudio.to('cpu')
waveform_tensor_torch = waveform_tensor_torch.to('cpu')
amplitude_to_DB = amplitude_to_DB.to('cpu')
feature_paddleaudio = benchmark(log_melspectrogram_torchaudio)
feature_librosa = librosa.feature.melspectrogram(waveform, **mel_conf)
feature_librosa = librosa.power_to_db(feature_librosa, top_db=80.0)
np.testing.assert_array_almost_equal(
feature_librosa, feature_paddleaudio, decimal=3)
def test_log_melspect_gpu_torchaudio(benchmark):
global waveform_tensor_torch, mel_extractor_torchaudio, amplitude_to_DB
mel_extractor_torchaudio = mel_extractor_torchaudio.to('cuda')
waveform_tensor_torch = waveform_tensor_torch.to('cuda')
amplitude_to_DB = amplitude_to_DB.to('cuda')
feature_torchaudio = benchmark(log_melspectrogram_torchaudio)
feature_librosa = librosa.feature.melspectrogram(waveform, **mel_conf)
feature_librosa = librosa.power_to_db(feature_librosa, top_db=80.0)
np.testing.assert_array_almost_equal(
feature_librosa, feature_torchaudio.cpu(), decimal=2)
paddleaudio/tests/benchmark/melspectrogram.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import urllib.request
import librosa
import numpy as np
import paddle
import torch
import torchaudio
import paddleaudio
wav_url = 'https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav'
if not os.path.isfile(os.path.basename(wav_url)):
urllib.request.urlretrieve(wav_url, os.path.basename(wav_url))
waveform, sr = paddleaudio.load(os.path.abspath(os.path.basename(wav_url)))
waveform_tensor = paddle.to_tensor(waveform).unsqueeze(0)
waveform_tensor_torch = torch.from_numpy(waveform).unsqueeze(0)
# Feature conf
mel_conf = {
'sr': sr,
'n_fft': 512,
'hop_length': 128,
'n_mels': 40,
}
mel_conf_torchaudio = {
'sample_rate': sr,
'n_fft': 512,
'hop_length': 128,
'n_mels': 40,
'norm': 'slaney',
'mel_scale': 'slaney',
}
def enable_cpu_device():
paddle.set_device('cpu')
def enable_gpu_device():
paddle.set_device('gpu')
mel_extractor = paddleaudio.features.MelSpectrogram(
**mel_conf, f_min=0.0, dtype=waveform_tensor.dtype)
def melspectrogram():
return mel_extractor(waveform_tensor).squeeze(0)
def test_melspect_cpu(benchmark):
enable_cpu_device()
feature_paddleaudio = benchmark(melspectrogram)
feature_librosa = librosa.feature.melspectrogram(waveform, **mel_conf)
np.testing.assert_array_almost_equal(
feature_librosa, feature_paddleaudio, decimal=3)
def test_melspect_gpu(benchmark):
enable_gpu_device()
feature_paddleaudio = benchmark(melspectrogram)
feature_librosa = librosa.feature.melspectrogram(waveform, **mel_conf)
np.testing.assert_array_almost_equal(
feature_librosa, feature_paddleaudio, decimal=3)
mel_extractor_torchaudio = torchaudio.transforms.MelSpectrogram(
**mel_conf_torchaudio, f_min=0.0)
def melspectrogram_torchaudio():
return mel_extractor_torchaudio(waveform_tensor_torch).squeeze(0)
def test_melspect_cpu_torchaudio(benchmark):
global waveform_tensor_torch, mel_extractor_torchaudio
mel_extractor_torchaudio = mel_extractor_torchaudio.to('cpu')
waveform_tensor_torch = waveform_tensor_torch.to('cpu')
feature_paddleaudio = benchmark(melspectrogram_torchaudio)
feature_librosa = librosa.feature.melspectrogram(waveform, **mel_conf)
np.testing.assert_array_almost_equal(
feature_librosa, feature_paddleaudio, decimal=3)
def test_melspect_gpu_torchaudio(benchmark):
global waveform_tensor_torch, mel_extractor_torchaudio
mel_extractor_torchaudio = mel_extractor_torchaudio.to('cuda')
waveform_tensor_torch = waveform_tensor_torch.to('cuda')
feature_torchaudio = benchmark(melspectrogram_torchaudio)
feature_librosa = librosa.feature.melspectrogram(waveform, **mel_conf)
np.testing.assert_array_almost_equal(
feature_librosa, feature_torchaudio.cpu(), decimal=3)
paddleaudio/tests/benchmark/mfcc.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import urllib.request
import librosa
import numpy as np
import paddle
import torch
import torchaudio
import paddleaudio
wav_url = 'https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav'
if not os.path.isfile(os.path.basename(wav_url)):
urllib.request.urlretrieve(wav_url, os.path.basename(wav_url))
waveform, sr = paddleaudio.load(os.path.abspath(os.path.basename(wav_url)))
waveform_tensor = paddle.to_tensor(waveform).unsqueeze(0)
waveform_tensor_torch = torch.from_numpy(waveform).unsqueeze(0)
# Feature conf
mel_conf = {
'sr': sr,
'n_fft': 512,
'hop_length': 128,
'n_mels': 40,
}
mfcc_conf = {
'n_mfcc': 20,
'top_db': 80.0,
}
mfcc_conf.update(mel_conf)
mel_conf_torchaudio = {
'sample_rate': sr,
'n_fft': 512,
'hop_length': 128,
'n_mels': 40,
'norm': 'slaney',
'mel_scale': 'slaney',
}
mfcc_conf_torchaudio = {
'sample_rate': sr,
'n_mfcc': 20,
}
def enable_cpu_device():
paddle.set_device('cpu')
def enable_gpu_device():
paddle.set_device('gpu')
mfcc_extractor = paddleaudio.features.MFCC(
**mfcc_conf, f_min=0.0, dtype=waveform_tensor.dtype)
def mfcc():
return mfcc_extractor(waveform_tensor).squeeze(0)
def test_mfcc_cpu(benchmark):
enable_cpu_device()
feature_paddleaudio = benchmark(mfcc)
feature_librosa = librosa.feature.mfcc(waveform, **mel_conf)
np.testing.assert_array_almost_equal(
feature_librosa, feature_paddleaudio, decimal=3)
def test_mfcc_gpu(benchmark):
enable_gpu_device()
feature_paddleaudio = benchmark(mfcc)
feature_librosa = librosa.feature.mfcc(waveform, **mel_conf)
np.testing.assert_array_almost_equal(
feature_librosa, feature_paddleaudio, decimal=3)
del mel_conf_torchaudio['sample_rate']
mfcc_extractor_torchaudio = torchaudio.transforms.MFCC(
**mfcc_conf_torchaudio, melkwargs=mel_conf_torchaudio)
def mfcc_torchaudio():
return mfcc_extractor_torchaudio(waveform_tensor_torch).squeeze(0)
def test_mfcc_cpu_torchaudio(benchmark):
global waveform_tensor_torch, mfcc_extractor_torchaudio
mel_extractor_torchaudio = mfcc_extractor_torchaudio.to('cpu')
waveform_tensor_torch = waveform_tensor_torch.to('cpu')
feature_paddleaudio = benchmark(mfcc_torchaudio)
feature_librosa = librosa.feature.mfcc(waveform, **mel_conf)
np.testing.assert_array_almost_equal(
feature_librosa, feature_paddleaudio, decimal=3)
def test_mfcc_gpu_torchaudio(benchmark):
global waveform_tensor_torch, mfcc_extractor_torchaudio
mel_extractor_torchaudio = mfcc_extractor_torchaudio.to('cuda')
waveform_tensor_torch = waveform_tensor_torch.to('cuda')
feature_torchaudio = benchmark(mfcc_torchaudio)
feature_librosa = librosa.feature.mfcc(waveform, **mel_conf)
np.testing.assert_array_almost_equal(
feature_librosa, feature_torchaudio.cpu(), decimal=3)
paddleaudio/tests/features/__init__.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
paddleaudio/tests/features/base.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import unittest
import urllib.request
import numpy as np
import paddle
from paddleaudio import load
wav_url = 'https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav'
class FeatTest(unittest.TestCase):
def setUp(self):
self.initParmas()
self.initWavInput()
self.setUpDevice()
def setUpDevice(self, device='cpu'):
paddle.set_device(device)
def initWavInput(self, url=wav_url):
if not os.path.isfile(os.path.basename(url)):
urllib.request.urlretrieve(url, os.path.basename(url))
self.waveform, self.sr = load(os.path.abspath(os.path.basename(url)))
self.waveform = self.waveform.astype(
np.float32
) # paddlespeech.s2t.transform.spectrogram only supports float32
dim = len(self.waveform.shape)
assert dim in [1, 2]
if dim == 1:
self.waveform = np.expand_dims(self.waveform, 0)
def initParmas(self):
raise NotImplementedError
paddleaudio/tests/features/test_istft.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import unittest
import numpy as np
import paddle
from .base import FeatTest
from paddleaudio.functional.window import get_window
from paddlespeech.s2t.transform.spectrogram import IStft
from paddlespeech.s2t.transform.spectrogram import Stft
class TestIstft(FeatTest):
def initParmas(self):
self.n_fft = 512
self.hop_length = 128
self.window_str = 'hann'
def test_istft(self):
ps_stft = Stft(self.n_fft, self.hop_length)
ps_res = ps_stft(
self.waveform.T).squeeze(1).T # (n_fft//2 + 1, n_frmaes)
x = paddle.to_tensor(ps_res)
ps_istft = IStft(self.hop_length)
ps_res = ps_istft(ps_res.T)
window = get_window(
self.window_str, self.n_fft, dtype=self.waveform.dtype)
pd_res = paddle.signal.istft(
x, self.n_fft, self.hop_length, window=window)
np.testing.assert_array_almost_equal(ps_res, pd_res, decimal=5)
if __name__ == '__main__':
unittest.main()
paddleaudio/tests/features/test_kaldi.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import unittest
import numpy as np
import paddle
import torch
import torchaudio
import paddleaudio
from .base import FeatTest
class TestKaldi(FeatTest):
def initParmas(self):
self.window_size = 1024
self.dtype = 'float32'
def test_window(self):
t_hann_window = torch.hann_window(
self.window_size, periodic=False, dtype=eval(f'torch.{self.dtype}'))
t_hamm_window = torch.hamming_window(
self.window_size,
periodic=False,
alpha=0.54,
beta=0.46,
dtype=eval(f'torch.{self.dtype}'))
t_povey_window = torch.hann_window(
self.window_size, periodic=False,
dtype=eval(f'torch.{self.dtype}')).pow(0.85)
p_hann_window = paddleaudio.functional.window.get_window(
'hann',
self.window_size,
fftbins=False,
dtype=eval(f'paddle.{self.dtype}'))
p_hamm_window = paddleaudio.functional.window.get_window(
'hamming',
self.window_size,
fftbins=False,
dtype=eval(f'paddle.{self.dtype}'))
p_povey_window = paddleaudio.functional.window.get_window(
'hann',
self.window_size,
fftbins=False,
dtype=eval(f'paddle.{self.dtype}')).pow(0.85)
np.testing.assert_array_almost_equal(t_hann_window, p_hann_window)
np.testing.assert_array_almost_equal(t_hamm_window, p_hamm_window)
np.testing.assert_array_almost_equal(t_povey_window, p_povey_window)
def test_fbank(self):
ta_features = torchaudio.compliance.kaldi.fbank(
torch.from_numpy(self.waveform.astype(self.dtype)))
pa_features = paddleaudio.compliance.kaldi.fbank(
paddle.to_tensor(self.waveform.astype(self.dtype)))
np.testing.assert_array_almost_equal(
ta_features, pa_features, decimal=4)
def test_mfcc(self):
ta_features = torchaudio.compliance.kaldi.mfcc(
torch.from_numpy(self.waveform.astype(self.dtype)))
pa_features = paddleaudio.compliance.kaldi.mfcc(
paddle.to_tensor(self.waveform.astype(self.dtype)))
np.testing.assert_array_almost_equal(
ta_features, pa_features, decimal=4)
if __name__ == '__main__':
unittest.main()
paddleaudio/tests/features/test_librosa.py 0 → 100644
浏览文件 @ c4035f8c
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import unittest
import librosa
import numpy as np
import paddle
import paddleaudio
from .base import FeatTest
from paddleaudio.functional.window import get_window
class TestLibrosa(FeatTest):
def initParmas(self):
self.n_fft = 512
self.hop_length = 128
self.n_mels = 40
self.n_mfcc = 20
self.fmin = 0.0
self.window_str = 'hann'
self.pad_mode = 'reflect'
self.top_db = 80.0
def test_stft(self):
if len(self.waveform.shape) == 2: # (C, T)
self.waveform = self.waveform.squeeze(
0) # 1D input for librosa.feature.melspectrogram
feature_librosa = librosa.core.stft(
y=self.waveform,
n_fft=self.n_fft,
hop_length=self.hop_length,
win_length=None,
window=self.window_str,
center=True,
dtype=None,
pad_mode=self.pad_mode, )
x = paddle.to_tensor(self.waveform).unsqueeze(0)
window = get_window(self.window_str, self.n_fft, dtype=x.dtype)
feature_paddle = paddle.signal.stft(
x=x,
n_fft=self.n_fft,
hop_length=self.hop_length,
win_length=None,
window=window,
center=True,
pad_mode=self.pad_mode,
normalized=False,
onesided=True, ).squeeze(0)
np.testing.assert_array_almost_equal(
feature_librosa, feature_paddle, decimal=5)
def test_istft(self):
if len(self.waveform.shape) == 2: # (C, T)
self.waveform = self.waveform.squeeze(
0) # 1D input for librosa.feature.melspectrogram
# Get stft result from librosa.
stft_matrix = librosa.core.stft(
y=self.waveform,
n_fft=self.n_fft,
hop_length=self.hop_length,
win_length=None,
window=self.window_str,
center=True,
pad_mode=self.pad_mode, )
feature_librosa = librosa.core.istft(
stft_matrix=stft_matrix,
hop_length=self.hop_length,
win_length=None,
window=self.window_str,
center=True,
dtype=None,
length=None, )
x = paddle.to_tensor(stft_matrix).unsqueeze(0)
window = get_window(
self.window_str,
self.n_fft,
dtype=paddle.to_tensor(self.waveform).dtype)
feature_paddle = paddle.signal.istft(
x=x,
n_fft=self.n_fft,
hop_length=self.hop_length,
win_length=None,
window=window,
center=True,
normalized=False,
onesided=True,
length=None,
return_complex=False, ).squeeze(0)
np.testing.assert_array_almost_equal(
feature_librosa, feature_paddle, decimal=5)
def test_mel(self):
feature_librosa = librosa.filters.mel(
sr=self.sr,
n_fft=self.n_fft,
n_mels=self.n_mels,
fmin=self.fmin,
fmax=None,
htk=False,
norm='slaney',
dtype=self.waveform.dtype, )
feature_compliance = paddleaudio.compliance.librosa.compute_fbank_matrix(
sr=self.sr,
n_fft=self.n_fft,
n_mels=self.n_mels,
fmin=self.fmin,
fmax=None,
htk=False,
norm='slaney',
dtype=self.waveform.dtype, )
x = paddle.to_tensor(self.waveform)
feature_functional = paddleaudio.functional.compute_fbank_matrix(
sr=self.sr,
n_fft=self.n_fft,
n_mels=self.n_mels,
f_min=self.fmin,
f_max=None,
htk=False,
norm='slaney',
dtype=x.dtype, )
np.testing.assert_array_almost_equal(feature_librosa,
feature_compliance)
np.testing.assert_array_almost_equal(feature_librosa,
feature_functional)
def test_melspect(self):
if len(self.waveform.shape) == 2: # (C, T)
self.waveform = self.waveform.squeeze(
0) # 1D input for librosa.feature.melspectrogram
# librosa:
feature_librosa = librosa.feature.melspectrogram(
y=self.waveform,
sr=self.sr,
n_fft=self.n_fft,
hop_length=self.hop_length,
n_mels=self.n_mels,
fmin=self.fmin)
# paddleaudio.compliance.librosa:
feature_compliance = paddleaudio.compliance.librosa.melspectrogram(
x=self.waveform,
sr=self.sr,
window_size=self.n_fft,
hop_length=self.hop_length,
n_mels=self.n_mels,
fmin=self.fmin,
to_db=False)
# paddleaudio.features.layer
x = paddle.to_tensor(
self.waveform, dtype=paddle.float64).unsqueeze(0) # Add batch dim.
feature_extractor = paddleaudio.features.MelSpectrogram(
sr=self.sr,
n_fft=self.n_fft,
hop_length=self.hop_length,
n_mels=self.n_mels,
f_min=self.fmin,
dtype=x.dtype)
feature_layer = feature_extractor(x).squeeze(0).numpy()
np.testing.assert_array_almost_equal(
feature_librosa, feature_compliance, decimal=5)
np.testing.assert_array_almost_equal(
feature_librosa, feature_layer, decimal=5)
def test_log_melspect(self):
if len(self.waveform.shape) == 2: # (C, T)
self.waveform = self.waveform.squeeze(
0) # 1D input for librosa.feature.melspectrogram
# librosa:
feature_librosa = librosa.feature.melspectrogram(
y=self.waveform,
sr=self.sr,
n_fft=self.n_fft,
hop_length=self.hop_length,
n_mels=self.n_mels,
fmin=self.fmin)
feature_librosa = librosa.power_to_db(feature_librosa, top_db=None)
# paddleaudio.compliance.librosa:
feature_compliance = paddleaudio.compliance.librosa.melspectrogram(
x=self.waveform,
sr=self.sr,
window_size=self.n_fft,
hop_length=self.hop_length,
n_mels=self.n_mels,
fmin=self.fmin)
# paddleaudio.features.layer
x = paddle.to_tensor(
self.waveform, dtype=paddle.float64).unsqueeze(0) # Add batch dim.
feature_extractor = paddleaudio.features.LogMelSpectrogram(
sr=self.sr,
n_fft=self.n_fft,
hop_length=self.hop_length,
n_mels=self.n_mels,
f_min=self.fmin,
dtype=x.dtype)
feature_layer = feature_extractor(x).squeeze(0).numpy()
np.testing.assert_array_almost_equal(
feature_librosa, feature_compliance, decimal=5)
np.testing.assert_array_almost_equal(
feature_librosa, feature_layer, decimal=4)
def test_mfcc(self):
if len(self.waveform.shape) == 2: # (C, T)
self.waveform = self.waveform.squeeze(
0) # 1D input for librosa.feature.melspectrogram
# librosa:
feature_librosa = librosa.feature.mfcc(
y=self.waveform,
sr=self.sr,
S=None,
n_mfcc=self.n_mfcc,
dct_type=2,
norm='ortho',
lifter=0,
n_fft=self.n_fft,
hop_length=self.hop_length,
n_mels=self.n_mels,
fmin=self.fmin)
# paddleaudio.compliance.librosa:
feature_compliance = paddleaudio.compliance.librosa.mfcc(
x=self.waveform,
sr=self.sr,
n_mfcc=self.n_mfcc,
dct_type=2,
norm='ortho',
lifter=0,
window_size=self.n_fft,
hop_length=self.hop_length,
n_mels=self.n_mels,
fmin=self.fmin,
top_db=self.top_db)
# paddleaudio.features.layer
x = paddle.to_tensor(
self.waveform, dtype=paddle.float64).unsqueeze(0) # Add batch dim.
feature_extractor = paddleaudio.features.MFCC(
sr=self.sr,
n_mfcc=self.n_mfcc,
n_fft=self.n_fft,
hop_length=self.hop_length,
n_mels=self.n_mels,
f_min=self.fmin,
top_db=self.top_db,
dtype=x.dtype)
feature_layer = feature_extractor(x).squeeze(0).numpy()
np.testing.assert_array_almost_equal(
feature_librosa, feature_compliance, decimal=4)
np.testing.assert_array_almost_equal(
feature_librosa, feature_layer, decimal=4)
if __name__ == '__main__':
unittest.main()
paddleaudio/tests/features/test_log_melspectrogram.py 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
paddleaudio/tests/features/test_spectrogram.py 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
paddleaudio/tests/features/test_stft.py 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
paddlespeech/__init__.py
浏览文件 @ c4035f8c
......@@ -11,3 +11,6 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import _locale
_locale._getdefaultlocale = (lambda *args: ['en_US', 'utf8'])
paddlespeech/cli/cls/infer.py
浏览文件 @ c4035f8c
......@@ -193,7 +193,8 @@ class CLSExecutor(BaseExecutor):
sr=feat_conf['sample_rate'],
mono=True,
dtype='float32')
logger.info("Preprocessing audio_file:" + audio_file)
if isinstance(audio_file, (str, os.PathLike)):
logger.info("Preprocessing audio_file:" + audio_file)
# Feature extraction
feature_extractor = LogMelSpectrogram(
......
paddlespeech/cli/utils.py
浏览文件 @ c4035f8c
......@@ -192,7 +192,7 @@ class ConfigCache:
try:
cfg = yaml.load(file, Loader=yaml.FullLoader)
self._data.update(cfg)
except:
except Exception as e:
self.flush()
@property
......
paddlespeech/server/__init__.py
浏览文件 @ c4035f8c
......@@ -18,6 +18,7 @@ from .base_commands import ClientHelpCommand
from .base_commands import ServerBaseCommand
from .base_commands import ServerHelpCommand
from .bin.paddlespeech_client import ASRClientExecutor
from .bin.paddlespeech_client import CLSClientExecutor
from .bin.paddlespeech_client import TTSClientExecutor
from .bin.paddlespeech_server import ServerExecutor
......
paddlespeech/server/bin/main.py
浏览文件 @ c4035f8c
此差异已折叠。
paddlespeech/server/conf/asr/asr.yaml 已删除 100644 → 0
浏览文件 @ 8d919781
此差异已折叠。
paddlespeech/server/conf/asr/asr_pd.yaml 已删除 100644 → 0
浏览文件 @ 8d919781
此差异已折叠。
paddlespeech/server/conf/tts/tts.yaml 已删除 100644 → 0
浏览文件 @ 8d919781
此差异已折叠。
paddlespeech/server/conf/tts/tts_pd.yaml 已删除 100644 → 0
浏览文件 @ 8d919781
此差异已折叠。
paddlespeech/server/engine/cls/__init__.py 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
paddlespeech/server/engine/cls/python/__init__.py 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
paddlespeech/server/engine/cls/python/cls_engine.py 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
paddlespeech/server/restful/api.py
浏览文件 @ c4035f8c
此差异已折叠。
paddlespeech/server/restful/cls_api.py 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
paddlespeech/t2s/exps/csmsc_test.txt 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
paddlespeech/t2s/exps/inference.py
浏览文件 @ c4035f8c
此差异已折叠。
paddlespeech/t2s/exps/syn_utils.py 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
paddlespeech/t2s/exps/synthesize.py
浏览文件 @ c4035f8c
此差异已折叠。
setup.py
浏览文件 @ c4035f8c
此差异已折叠。
speechx/.gitignore 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/CMakeLists.txt
浏览文件 @ c4035f8c
此差异已折叠。
speechx/README.md 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/build.sh 0 → 100755
浏览文件 @ c4035f8c
此差异已折叠。
speechx/cmake/EnableCMP0048.cmake 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/cmake/external/absl.cmake 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/cmake/external/boost.cmake 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/cmake/external/eigen.cmake 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/cmake/external/gflags.cmake 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/cmake/external/glog.cmake 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/cmake/external/gtest.cmake 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/cmake/external/kenlm.cmake 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/cmake/external/libsndfile.cmake 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/cmake/external/openblas.cmake 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/cmake/external/openfst.cmake 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/.gitignore 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/CMakeLists.txt 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/README.md 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/decoder/CMakeLists.txt 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/decoder/offline_decoder_main.cc 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/decoder/path.sh 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/decoder/run.sh 0 → 100755
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/decoder/valgrind.sh 0 → 100755
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/feat/CMakeLists.txt 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/feat/feature-mfcc-test.cc 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/feat/linear_spectrogram_main.cc 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/feat/path.sh 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/feat/run.sh 0 → 100755
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/feat/valgrind.sh 0 → 100755
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/nnet/CMakeLists.txt 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/nnet/path.sh 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/nnet/pp-model-test.cc 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/nnet/run.sh 0 → 100755
浏览文件 @ c4035f8c
此差异已折叠。
speechx/examples/nnet/valgrind.sh 0 → 100755
浏览文件 @ c4035f8c
此差异已折叠。
speechx/patch/CPPLINT.cfg 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/patch/openfst/src/include/fst/flags.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/patch/openfst/src/include/fst/log.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/patch/openfst/src/lib/flags.cc 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/CMakeLists.txt
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/base/basic_types.h
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/base/common.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/base/flags.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/base/log.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/base/macros.h
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/base/thread_pool.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/codelab/README.md 已删除 100644 → 0
浏览文件 @ 8d919781
此差异已折叠。
speechx/speechx/decoder/common.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/decoder/ctc_beam_search_decoder.cc 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/decoder/ctc_beam_search_decoder.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/decoder/ctc_decoders 0 → 120000
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/frontend/fbank.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/frontend/feature_cache.cc 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/frontend/feature_cache.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/frontend/feature_extractor_controller.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/frontend/feature_extractor_interface.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/frontend/linear_spectrogram.cc 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/frontend/linear_spectrogram.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/frontend/mfcc.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/frontend/normalizer.cc 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/frontend/normalizer.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/frontend/raw_audio.cc 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/frontend/raw_audio.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/frontend/window.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/kaldi/decoder/lattice-faster-decoder.cc 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/kaldi/decoder/lattice-faster-decoder.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/kaldi/feat/cmvn.cc 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/kaldi/feat/cmvn.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/kaldi/lat/determinize-lattice-pruned.cc 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/kaldi/lat/determinize-lattice-pruned.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/kaldi/lat/kaldi-lattice.cc 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/kaldi/lat/kaldi-lattice.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/kaldi/lat/lattice-functions.cc 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/kaldi/lat/lattice-functions.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/kaldi/matrix/BUILD 已删除 100644 → 0
浏览文件 @ 8d919781
此差异已折叠。
speechx/speechx/nnet/CMakeLists.txt 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/nnet/decodable-itf.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/nnet/decodable.cc 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/nnet/decodable.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/nnet/nnet_interface.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/nnet/paddle_nnet.cc 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/nnet/paddle_nnet.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/utils/file_utils.cc 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/speechx/utils/file_utils.h 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
speechx/tools/setup_valgrind.sh 0 → 100755
浏览文件 @ c4035f8c
此差异已折叠。
tests/test_tipc/benchmark_train.sh
浏览文件 @ c4035f8c
此差异已折叠。
tests/test_tipc/prepare.sh
浏览文件 @ c4035f8c
此差异已折叠。
tests/unit/cli/test_cli.sh
浏览文件 @ c4035f8c
此差异已折叠。
tests/unit/server/change_yaml.py 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
tests/unit/server/conf/application.yaml 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
tests/unit/server/test_server_client.sh 0 → 100644
浏览文件 @ c4035f8c
此差异已折叠。
utils/generate_infer_yaml.py
浏览文件 @ c4035f8c
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