This example contains code used to train a [Tacotron2 ](https://arxiv.org/abs/1712.05884) model with [AISHELL-3](http://www.aishelltech.com/aishell_3). The trained model can be used in Voice Cloning Task, We refer to the model structure of [Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis](https://arxiv.org/pdf/1806.04558.pdf) . The general steps are as follows:
This example contains code used to train a [Tacotron2 ](https://arxiv.org/abs/1712.05884) model with [AISHELL-3](http://www.aishelltech.com/aishell_3). The trained model can be used in Voice Cloning Task, We refer to the model structure of [Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis](https://arxiv.org/pdf/1806.04558.pdf) . The general steps are as follows:
1. Speaker Encoder: We use a Speaker Verification to train a speaker encoder. Datasets used in this task are different from those used in Tacotron2, because the transcriptions are not needed, we use more datasets, refer to [ge2e](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/ge2e).
1. Speaker Encoder: We use a Speaker Verification to train a speaker encoder. Datasets used in this task are different from those used in Tacotron2, because the transcriptions are not needed, we use more datasets, refer to [ge2e](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/ge2e).
2. Synthesizer: Then, we use the trained speaker encoder to generate utterance embedding for each sentence in AISHELL-3. This embedding is a extra input of Tacotron2 which will be concated with encoder outputs.
2. Synthesizer: We use the trained speaker encoder to generate speaker embedding for each sentence in AISHELL-3. This embedding is an extra input of Tacotron2 which will be concated with encoder outputs.
3. Vocoder: We use WaveFlow as the neural Vocoder, refer to [waveflow](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/ljspeech/voc0).
3. Vocoder: We use WaveFlow as the neural Vocoder, refer to [waveflow](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/ljspeech/voc0).
## Get Started
## Get Started
...
@@ -10,7 +10,7 @@ Assume the path to the MFA result of AISHELL-3 is `./alignment`.
...
@@ -10,7 +10,7 @@ Assume the path to the MFA result of AISHELL-3 is `./alignment`.
Assume the path to the pretrained ge2e model is `ge2e_ckpt_path=./ge2e_ckpt_0.3/step-3000000`
Assume the path to the pretrained ge2e model is `ge2e_ckpt_path=./ge2e_ckpt_0.3/step-3000000`
Use pretrained GE2E (speaker encoder) to generate utterance embedding for each sentence in AISHELL-3, which has the same file structure with wav files and the format is `.npy`.
Use pretrained GE2E (speaker encoder) to generate speaker embedding for each sentence in AISHELL-3, which has the same file structure with wav files and the format is `.npy`.
```bash
```bash
if[${stage}-le 0 ]&&[${stop_stage}-ge 0 ];then
if[${stage}-le 0 ]&&[${stop_stage}-ge 0 ];then
...
@@ -33,8 +33,8 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
...
@@ -33,8 +33,8 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
fi
fi
```
```
The computing time of utterance embedding can be x hours.
The computing time of utterance embedding can be x hours.
#### process wav
#### process wav
There are silence in the edge of AISHELL-3's wavs, and the audio amplitude is very small, so, we need to remove the silence and normalize the audio. You can the silence remove method based on volume or energy, but the effect is not very good, We use [MFA](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) to get the alignment of text and speech, then utilize the alignment results to remove the silence.
There are silence in the edge of AISHELL-3's wavs, and the audio amplitude is very small, so, we need to remove the silence and normalize the audio. You can the silence remove method based on volume or energy, but the effect is not very good, We use [MFA](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) to get the alignment of text and speech, then utilize the alignment results to remove the silence.
We use Montreal Force Aligner 1.0. The label in aishell3 include pinyin，so the lexicon we provided to MFA is pinyin rather than Chinese characters. And the prosody marks(`$` and `%`) need to be removed. You shoud preprocess the dataset into the format which MFA needs, the texts have the same name with wavs and have the suffix `.lab`.
We use Montreal Force Aligner 1.0. The label in aishell3 include pinyin，so the lexicon we provided to MFA is pinyin rather than Chinese characters. And the prosody marks(`$` and `%`) need to be removed. You shoud preprocess the dataset into the format which MFA needs, the texts have the same name with wavs and have the suffix `.lab`.
...
@@ -73,7 +73,7 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
...
@@ -73,7 +73,7 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
Our model remve stop token prediction in Tacotron2, because of the problem of extremely unbalanced proportion of positive and negative samples of stop token prediction, and it's very sensitive to the clip of audio silence. We use the last symbol from the highest point of attention to the encoder side as the termination condition.
Our model remve stop token prediction in Tacotron2, because of the problem of extremely unbalanced proportion of positive and negative samples of stop token prediction, and it's very sensitive to the clip of audio silence. We use the last symbol from the highest point of attention to the encoder side as the termination condition.
In addition, in order to accelerate the convergence of the model, we add `guided attention loss` to induce the alignment between encoder and decoder to show diagonal lines faster.
In addition, in order to accelerate the convergence of the model, we add `guided attention loss` to induce the alignment between encoder and decoder to show diagonal lines faster.
This example contains code used to train a [Tacotron2 ](https://arxiv.org/abs/1712.05884) model with [AISHELL-3](http://www.aishelltech.com/aishell_3). The trained model can be used in Voice Cloning Task, We refer to the model structure of [Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis](https://arxiv.org/pdf/1806.04558.pdf) . The general steps are as follows:
This example contains code used to train a [FastSpeech2](https://arxiv.org/abs/2006.04558) model with [AISHELL-3](http://www.aishelltech.com/aishell_3). The trained model can be used in Voice Cloning Task, We refer to the model structure of [Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis](https://arxiv.org/pdf/1806.04558.pdf) . The general steps are as follows:
1. Speaker Encoder: We use a Speaker Verification to train a speaker encoder. Datasets used in this task are different from those used in Tacotron2, because the transcriptions are not needed, we use more datasets, refer to [ge2e](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/ge2e).
1. Speaker Encoder: We use a Speaker Verification to train a speaker encoder. Datasets used in this task are different from those used in `FastSpeech2`, because the transcriptions are not needed, we use more datasets, refer to [ge2e](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/ge2e).
2. Synthesizer: Then, we use the trained speaker encoder to generate utterance embedding for each sentence in AISHELL-3. This embedding is a extra input of Tacotron2 which will be concated with encoder outputs.
2. Synthesizer: We use the trained speaker encoder to generate speaker embedding for each sentence in AISHELL-3. This embedding is an extra input of `FastSpeech2` which will be concated with encoder outputs.
3. Vocoder: We use WaveFlow as the neural Vocoder, refer to [waveflow](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/ljspeech/voc0).
3. Vocoder: We use [Parallel Wave GAN](http://arxiv.org/abs/1910.11480) as the neural Vocoder, refer to [voc1](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/aishell3/voc1).
We use [MFA2.x](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) to get durations for aishell3_fastspeech2.
You can download from here [aishell3_alignment_tone.tar.gz](https://paddlespeech.bj.bcebos.com/MFA/AISHELL-3/with_tone/aishell3_alignment_tone.tar.gz), or train your own MFA model reference to [use_mfa example](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/use_mfa)(use MFA1.x now) of our repo.
## Pretrained GE2E model
We use pretrained GE2E model to generate spwaker embedding for each sentence.
Download pretrained GE2E model from here [ge2e_ckpt_0.3.zip](https://paddlespeech.bj.bcebos.com/Parakeet/ge2e_ckpt_0.3.zip), and `unzip` it.
## Get Started
## Get Started
Assume the path to the dataset is `~/datasets/data_aishell3`.
Assume the path to the dataset is `~/datasets/data_aishell3`.
Assume the path to the MFA result of AISHELL-3 is `./alignment`.
Assume the path to the MFA result of AISHELL-3 is `./aishell3_alignment_tone`.
Assume the path to the pretrained ge2e model is `ge2e_ckpt_path=./ge2e_ckpt_0.3/step-3000000`
Assume the path to the pretrained ge2e model is `./ge2e_ckpt_0.3`.
When it is done. A `dump` folder is created in the current directory. The structure of the dump folder is listed below.
Use pretrained GE2E (speaker encoder) to generate utterance embedding for each sentence in AISHELL-3, which has the same file structure with wav files and the format is `.npy`.
```text
dump
```bash
├── dev
if[${stage}-le 0 ]&&[${stop_stage}-ge 0 ];then
│ ├── norm
python3 ${BIN_DIR}/../ge2e/inference.py \
│ └── raw
--input=${input}\
├── embed
--output=${preprocess_path}/embed \
│ ├── SSB0005
--ngpu=1 \
│ ├── SSB0009
--checkpoint_path=${ge2e_ckpt_path}
│ ├── ...
fi
│ └── ...
├── phone_id_map.txt
├── speaker_id_map.txt
├── test
│ ├── norm
│ └── raw
└── train
├── energy_stats.npy
├── norm
├── pitch_stats.npy
├── raw
└── speech_stats.npy
```
```
The `embed` contains the generated speaker embedding for each sentence in AISHELL-3, which has the same file structure with wav files and the format is `.npy`.
The computing time of utterance embedding can be x hours.
The computing time of utterance embedding can be x hours.
#### process wav
There are silence in the edge of AISHELL-3's wavs, and the audio amplitude is very small, so, we need to remove the silence and normalize the audio. You can the silence remove method based on volume or energy, but the effect is not very good, We use [MFA](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) to get the alignment of text and speech, then utilize the alignment results to remove the silence.
We use Montreal Force Aligner 1.0. The label in aishell3 include pinyin，so the lexicon we provided to MFA is pinyin rather than Chinese characters. And the prosody marks(`$` and `%`) need to be removed. You shoud preprocess the dataset into the format which MFA needs, the texts have the same name with wavs and have the suffix `.lab`.
The dataset is split into 3 parts, namely `train`, `dev` and` test`, each of which contains a `norm` and `raw` sub folder. The raw folder contains speech、pitch and energy features of each utterances, while the norm folder contains normalized ones. The statistics used to normalize features are computed from the training set, which is located in `dump/train/*_stats.npy`.
We use [lexicon.txt](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/paddlespeech/t2s/exps/voice_cloning/tacotron2_ge2e/lexicon.txt) as the lexicon.
Also there is a `metadata.jsonl` in each subfolder. It is a table-like file which contains phones, text_lengths, speech_lengths, durations, path of speech features, path of pitch features, path of energy features, speaker and id of each utterance.
You can download the alignment results from here [alignment_aishell3.tar.gz](https://paddlespeech.bj.bcebos.com/Parakeet/alignment_aishell3.tar.gz), or train your own MFA model reference to [use_mfa example](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/use_mfa)(use MFA1.x now) of our repo.
The preprocessing step is very similar to that one of [tts3](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/aishell3/tts3), but there is one more `ge2e/inference` step here.
The training step is very similar to that one of [tts3](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/aishell3/tts3), but we should set `--voice-cloning=True` when calling `${BIN_DIR}/train.py`.
#### preprocess transcription
### Synthesize
We revert the transcription into `phones` and `tones`. It is worth noting that our processing here is different from that used for MFA, we separated the tones. This is a processing method, of course, you can only segment initials and vowels.
We use [parallel wavegan](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/aishell3/voc1) as the neural vocoder.
Download pretrained parallel wavegan model from [pwg_aishell3_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/pwg_aishell3_ckpt_0.5.zip) and unzip it.
```bash
```bash
if[${stage}-le 2 ]&&[${stop_stage}-ge 2 ];then
unzip pwg_aishell3_ckpt_0.5.zip
python3 ${BIN_DIR}/preprocess_transcription.py \
--input=${input}\
--output=${preprocess_path}
fi
```
```
The default input is `~/datasets/data_aishell3/train`，which contains `label_train-set.txt`, the processed results are `metadata.yaml` and `metadata.pickle`. the former is a text format for easy viewing, and the latter is a binary format for direct reading.
The synthesizing step is very similar to that one of [tts3](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/aishell3/tts3), but we should set `--voice-cloning=True` when calling `${BIN_DIR}/synthesize.py`.
Our model remve stop token prediction in Tacotron2, because of the problem of extremely unbalanced proportion of positive and negative samples of stop token prediction, and it's very sensitive to the clip of audio silence. We use the last symbol from the highest point of attention to the encoder side as the termination condition.
### Voice Cloning
Assume there are some reference audios in `./ref_audio`
In addition, in order to accelerate the convergence of the model, we add `guided attention loss` to induce the alignment between encoder and decoder to show diagonal lines faster.
Pretrained SpeedySpeech model with no silence in the edge of audios. [speedyspeech_nosil_baker_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/speedyspeech_nosil_baker_ckpt_0.5.zip)
Pretrained SpeedySpeech model with no silence in the edge of audios[speedyspeech_nosil_baker_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/speedyspeech_nosil_baker_ckpt_0.5.zip).
Static model can be downloaded here [speedyspeech_nosil_baker_static_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/speedyspeech_nosil_baker_static_0.5.zip).
Static model can be downloaded here [speedyspeech_nosil_baker_static_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/speedyspeech_nosil_baker_static_0.5.zip).
Pretrained FastSpeech2 model with no silence in the edge of audios. [fastspeech2_nosil_baker_ckpt_0.4.zip](https://paddlespeech.bj.bcebos.com/Parakeet/fastspeech2_nosil_baker_ckpt_0.4.zip)
Pretrained FastSpeech2 model with no silence in the edge of audios [fastspeech2_nosil_baker_ckpt_0.4.zip](https://paddlespeech.bj.bcebos.com/Parakeet/fastspeech2_nosil_baker_ckpt_0.4.zip).
Static model can be downloaded here [fastspeech2_nosil_baker_static_0.4.zip](https://paddlespeech.bj.bcebos.com/Parakeet/fastspeech2_nosil_baker_static_0.4.zip)
Static model can be downloaded here [fastspeech2_nosil_baker_static_0.4.zip](https://paddlespeech.bj.bcebos.com/Parakeet/fastspeech2_nosil_baker_static_0.4.zip).
@@ -13,7 +13,7 @@ Assume the path to the dataset is `~/datasets/BZNSYP`.
...
@@ -13,7 +13,7 @@ Assume the path to the dataset is `~/datasets/BZNSYP`.
Assume the path to the MFA result of CSMSC is `./baker_alignment_tone`.
Assume the path to the MFA result of CSMSC is `./baker_alignment_tone`.
Run the command below to
Run the command below to
1.**source path**.
1.**source path**.
2. preprocess the dataset,
2. preprocess the dataset.
3. train the model.
3. train the model.
4. synthesize wavs.
4. synthesize wavs.
- synthesize waveform from `metadata.jsonl`.
- synthesize waveform from `metadata.jsonl`.
...
@@ -106,8 +106,51 @@ optional arguments:
...
@@ -106,8 +106,51 @@ optional arguments:
4.`--output-dir` is the directory to save the synthesized audio files.
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.
5.`--ngpu` is the number of gpus to use, if ngpu == 0, use cpu.
## Finetune
Since there are no `noise` in the input of Multi Band MelGAN, the audio quality is not so good (see [espnet issue](https://github.com/espnet/espnet/issues/3536#issuecomment-916035415)), we refer to the method proposed in [HiFiGAN](https://arxiv.org/abs/2010.05646), finetune Multi Band MelGAN with the predicted mel-spectrogram from `FastSpeech2`.
The length of mel-spectrograms should align with the length of wavs, so we should generate mels using ground truth alignment.
But since we are fine-tuning, we should use the statistics computed during training step.
You should first download pretrained `FastSpeech2` model from [fastspeech2_nosil_baker_ckpt_0.4.zip](https://paddlespeech.bj.bcebos.com/Parakeet/fastspeech2_nosil_baker_ckpt_0.4.zip) and `unzip` it.
Assume the path to the dump-dir of training step is `dump`.
Assume the path to the duration result of CSMSC is `durations.txt` (generated during training step's preprocessing).
Assume the path to the pretrained `FastSpeech2` model is `fastspeech2_nosil_baker_ckpt_0.4`.
\
The `finetune.sh` can
1.**source path**.
2. generate ground truth alignment mels.
3. link `*_wave.npy` from `dump` to `dump_finetune` (because we only use new mels, the wavs are the ones used during train step) .
4. copy features' stats from `dump` to `dump_finetune`.
5. normalize the ground truth alignment mels.
6. finetune the model.
Before finetune, make sure that the pretrained model is in `finetune.sh` 's `${output-dir}/checkpoints`, and there is a `records.jsonl` in it to refer to this pretrained model
```text
exp/finetune/checkpoints
├── records.jsonl
└── snapshot_iter_1000000.pdz
```
The content of `records.jsonl` should be as follows (change `"path"` to your own ckpt path):
By default, `finetune.sh` will use `conf/finetune.yaml` as config, the dump-dir is `dump_finetune`, the experiment dir is `exp/finetune`.
TODO:
The hyperparameter of `finetune.yaml` is not good enough, a smaller `learning_rate` should be used (more `milestones` should be set).
## Pretrained Models
## Pretrained Models
Pretrained model can be downloaded here [mb_melgan_baker_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/mb_melgan_baker_ckpt_0.5.zip).
Pretrained model can be downloaded here [mb_melgan_baker_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/mb_melgan_baker_ckpt_0.5.zip).
Finetuned model can ben downloaded here [mb_melgan_baker_finetune_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/mb_melgan_baker_finetune_ckpt_0.5.zip).
Static model can be downloaded here [mb_melgan_baker_static_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/mb_melgan_baker_static_0.5.zip)
Static model can be downloaded here [mb_melgan_baker_static_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/mb_melgan_baker_static_0.5.zip)
Multi Band MelGAN checkpoint contains files listed below.
Multi Band MelGAN checkpoint contains files listed below.
