14 KiB

Raw Blame History

Finetune your own AM based on FastSpeech2 with multi-speakers dataset.

This example shows how to finetune your own AM based on FastSpeech2 with multi-speakers dataset. For finetuning Chinese data, we use part of csmsc's data (top 200) and Fastspeech2 pretrained model with AISHELL-3. For finetuning English data, we use part of ljspeech's data (top 200) and Fastspeech2 pretrained model with VCTK. The example is implemented according to this discussion. Thanks to the developer for the idea.

For more information on training Fastspeech2 with AISHELL-3, You can refer examples/aishell3/tts3. For more information on training Fastspeech2 with VCTK, You can refer examples/vctk/tts3.

Prepare

Download Pretrained model

Assume the path to the model is ./pretrained_models.
If you want to finetune Chinese pretrained model, you need to download Fastspeech2 pretrained model with AISHELL-3: fastspeech2_aishell3_ckpt_1.1.0.zip for finetuning. Download HiFiGAN pretrained model with aishell3: hifigan_aishell3_ckpt_0.2.0.zip for synthesis.

mkdir -p pretrained_models && cd pretrained_models
# pretrained fastspeech2 model
wget https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_aishell3_ckpt_1.1.0.zip 
unzip fastspeech2_aishell3_ckpt_1.1.0.zip
# pretrained hifigan model
wget https://paddlespeech.bj.bcebos.com/Parakeet/released_models/hifigan/hifigan_aishell3_ckpt_0.2.0.zip
unzip hifigan_aishell3_ckpt_0.2.0.zip
cd ../

If you want to finetune English pretrained model, you need to download Fastspeech2 pretrained model with VCTK: fastspeech2_vctk_ckpt_1.2.0.zip for finetuning. Download HiFiGAN pretrained model with VCTK: hifigan_vctk_ckpt_0.2.0.zip for synthesis.

mkdir -p pretrained_models && cd pretrained_models
# pretrained fastspeech2 model
wget https://paddlespeech.bj.bcebos.com/Parakeet/released_models/fastspeech2/fastspeech2_vctk_ckpt_1.2.0.zip 
unzip fastspeech2_vctk_ckpt_1.2.0.zip
# pretrained hifigan model
wget https://paddlespeech.bj.bcebos.com/Parakeet/released_models/hifigan/hifigan_vctk_ckpt_0.2.0.zip
unzip hifigan_vctk_ckpt_0.2.0.zip
cd ../

If you want to finetune Chinese-English Mixed pretrained model, you need to download Fastspeech2 pretrained model with mix datasets: fastspeech2_mix_ckpt_1.2.0.zip for finetuning. Download HiFiGAN pretrained model with aishell3: hifigan_aishell3_ckpt_0.2.0.zip for synthesis.

mkdir -p pretrained_models && cd pretrained_models
# pretrained fastspeech2 model
wget https://paddlespeech.bj.bcebos.com/t2s/chinse_english_mixed/models/fastspeech2_mix_ckpt_1.2.0.zip
unzip fastspeech2_mix_ckpt_1.2.0.zip
# pretrained hifigan model
wget https://paddlespeech.bj.bcebos.com/Parakeet/released_models/hifigan/hifigan_aishell3_ckpt_0.2.0.zip
unzip hifigan_aishell3_ckpt_0.2.0.zip
cd ../

Prepare your data

Assume the path to the dataset is ./input which contains a speaker folder. Speaker folder contains audio files (*.wav) and label file (labels.txt). The format of the audio file is wav. The format of the label file is: utt_id|pronunciation.

If you want to finetune Chinese pretrained model, you need to prepare Chinese data. Chinese label example:

000001|ka2 er2 pu3 pei2 wai4 sun1 wan2 hua2 ti1

Here is an example of the first 200 data of csmsc.

mkdir -p input && cd input
wget https://paddlespeech.bj.bcebos.com/datasets/csmsc_mini.zip
unzip csmsc_mini.zip
cd ../

If you want to finetune English pretrained model, you need to prepare English data. English label example:

LJ001-0001|Printing, in the only sense with which we are at present concerned, differs from most if not from all the arts and crafts represented in the Exhibition

Here is an example of the first 200 data of ljspeech.

mkdir -p input && cd input
wget https://paddlespeech.bj.bcebos.com/datasets/ljspeech_mini.zip
unzip ljspeech_mini.zip
cd ../

If you want to finetune Chinese-English Mixed pretrained model, you need to prepare Chinese data or English data. Here is an example of the first 12 data of SSB0005 (the speaker of aishell3).

mkdir -p input && cd input
wget https://paddlespeech.bj.bcebos.com/datasets/SSB0005_mini.zip
unzip SSB0005_mini.zip
cd ../

Download MFA tools and pretrained model

Assume the path to the MFA tool is ./tools. Download MFA.

mkdir -p tools && cd tools
# mfa tool
wget https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner/releases/download/v1.0.1/montreal-forced-aligner_linux.tar.gz
tar xvf montreal-forced-aligner_linux.tar.gz
cp montreal-forced-aligner/lib/libpython3.6m.so.1.0 montreal-forced-aligner/lib/libpython3.6m.so
mkdir -p aligner && cd aligner

If you want to get mfa result of Chinese data, you need to download pretrained MFA models with aishell3: aishell3_model.zip and unzip it.

# pretrained mfa model for Chinese data
wget https://paddlespeech.bj.bcebos.com/MFA/ernie_sat/aishell3_model.zip
unzip aishell3_model.zip
wget https://paddlespeech.bj.bcebos.com/MFA/AISHELL-3/with_tone/simple.lexicon
cd ../../

If you want to get mfa result of English data, you need to download pretrained MFA models with vctk: vctk_model.zip and unzip it.

# pretrained mfa model for English data
wget https://paddlespeech.bj.bcebos.com/MFA/ernie_sat/vctk_model.zip
unzip vctk_model.zip
wget https://paddlespeech.bj.bcebos.com/MFA/LJSpeech-1.1/cmudict-0.7b
cd ../../

When "Prepare" done. The structure of the current directory is similar to the following.

├── input
│   ├── csmsc_mini
│   │   ├── 000001.wav
│   │   ├── 000002.wav
│   │   ├── 000003.wav
│   │   ├── ...
│   │   ├── 000200.wav
│   │   ├── labels.txt
│   └── csmsc_mini.zip
├── pretrained_models
│   ├── fastspeech2_aishell3_ckpt_1.1.0
│   │   ├── default.yaml
│   │   ├── energy_stats.npy
│   │   ├── phone_id_map.txt
│   │   ├── pitch_stats.npy
│   │   ├── snapshot_iter_96400.pdz
│   │   ├── speaker_id_map.txt
│   │   └── speech_stats.npy
│   ├── fastspeech2_aishell3_ckpt_1.1.0.zip
│   ├── hifigan_aishell3_ckpt_0.2.0    
│   │   ├── default.yaml
│   │   ├── feats_stats.npy
│   │   └── snapshot_iter_2500000.pdz
│   └── hifigan_aishell3_ckpt_0.2.0.zip
└── tools
    ├── aligner
    │   ├── aishell3_model
    │   ├── aishell3_model.zip
    │   └── simple.lexicon
    ├── montreal-forced-aligner
    │   ├── bin
    │   ├── lib
    │   └── pretrained_models
    └── montreal-forced-aligner_linux.tar.gz
    ...

Set finetune.yaml

conf/finetune.yaml contains some configurations for fine-tuning. You can try various options to fine better result. The value of frozen_layers can be change according conf/fastspeech2_layers.txt which is the model layer of fastspeech2.

Arguments:

batch_size: finetune batch size which should be less than or equal to the number of training samples. Default: -1, means 64 which same to pretrained model
learning_rate: learning rate. Default: 0.0001
num_snapshots: number of save models. Default: -1, means 5 which same to pretrained model
frozen_layers: frozen layers. must be a list. If you don't want to frozen any layer, set [].

Get Started

For finetuning Chinese pretrained model, execute ./run.sh. For finetuning English pretrained model, execute ./run_en.sh. For finetuning Chinese-English Mixed pretrained model, execute ./run_mix.sh.
Run the command below to

source path.
finetune the model.
synthesize wavs.
- synthesize waveform from text file.

./run.sh

You can choose a range of stages you want to run, or set stage equal to stop-stage to run only one stage.

Model Finetune

Finetune a FastSpeech2 model.

./run.sh --stage 0 --stop-stage 5

stage 5 of run.sh calls local/finetune.py, here's the complete help message.

usage: finetune.py [-h] [--pretrained_model_dir PRETRAINED_MODEL_DIR]
                [--dump_dir DUMP_DIR] [--output_dir OUTPUT_DIR] [--ngpu NGPU]
                [--epoch EPOCH] [--finetune_config FINETUNE_CONFIG]

optional arguments:
  -h, --help           Show this help message and exit
  --pretrained_model_dir PRETRAINED_MODEL_DIR
                       Path to pretrained model
  --dump_dir DUMP_DIR
                       directory to save feature files and metadata
  --output_dir OUTPUT_DIR      
                       Directory to save finetune model 
  --ngpu NGPU          The number of gpu, if ngpu=0, use cpu
  --epoch EPOCH        The epoch of finetune
  --finetune_config FINETUNE_CONFIG        
                       Path to finetune config file

--pretrained_model_dir is the directory incluing pretrained fastspeech2_aishell3 model.
--dump_dir is the directory including audio feature and metadata.
--output_dir is the directory to save finetune model.
--ngpu is the number of gpu, if ngpu=0, use cpu
--epoch is the epoch of finetune.
--finetune_config is the path to finetune config file

Synthesizing

To synthesize Chinese audio, We use HiFiGAN with aishell3 as the neural vocoder. Assume the path to the hifigan model is ./pretrained_models. Download the pretrained HiFiGAN model from hifigan_aishell3_ckpt_0.2.0 and unzip it.

To synthesize English audio, We use HiFiGAN with vctk as the neural vocoder. Assume the path to the hifigan model is ./pretrained_models. Download the pretrained HiFiGAN model from hifigan_vctk_ckpt_0.2.0.zip and unzip it.

Modify ckpt in run.sh to the final model in exp/default/checkpoints.

./run.sh --stage 6 --stop-stage 6

stage 6 of run.sh calls ${BIN_DIR}/../synthesize_e2e.py, which can synthesize waveform from text file.

usage: synthesize_e2e.py [-h]
                         [--am {fastspeech2_aishell3,fastspeech2_vctk}]
                         [--am_config AM_CONFIG] [--am_ckpt AM_CKPT]
                         [--am_stat AM_STAT] [--phones_dict PHONES_DICT]
                         [--tones_dict TONES_DICT]
                         [--speaker_dict SPEAKER_DICT] [--spk_id SPK_ID]
                         [--voc {pwgan_aishell3, pwgan_vctk, hifigan_aishell3, hifigan_vctk}]
                         [--voc_config VOC_CONFIG] [--voc_ckpt VOC_CKPT]
                         [--voc_stat VOC_STAT] [--lang LANG]
                         [--inference_dir INFERENCE_DIR] [--ngpu NGPU]
                         [--text TEXT] [--output_dir OUTPUT_DIR]

Synthesize with acoustic model & vocoder

optional arguments:
  -h, --help            show this help message and exit
  --am {fastspeech2_aishell3, fastspeech2_vctk}
                        Choose acoustic model type of tts task.
  --am_config AM_CONFIG
                        Config of acoustic model.
  --am_ckpt AM_CKPT     Checkpoint file of acoustic model.
  --am_stat AM_STAT     mean and standard deviation used to normalize
                        spectrogram when training acoustic model.
  --phones_dict PHONES_DICT
                        phone vocabulary file.
  --tones_dict TONES_DICT
                        tone vocabulary file.
  --speaker_dict SPEAKER_DICT
                        speaker id map file.
  --spk_id SPK_ID       spk id for multi speaker acoustic model
  --voc {pwgan_aishell3, pwgan_vctk, hifigan_aishell3, hifigan_vctk}
                        Choose vocoder type of tts task.
  --voc_config VOC_CONFIG
                        Config of voc.
  --voc_ckpt VOC_CKPT   Checkpoint file of voc.
  --voc_stat VOC_STAT   mean and standard deviation used to normalize
                        spectrogram when training voc.
  --lang LANG           Choose model language. zh or en
  --inference_dir INFERENCE_DIR
                        dir to save inference models
  --ngpu NGPU           if ngpu == 0, use cpu.
  --text TEXT           text to synthesize, a 'utt_id sentence' pair per line.
  --output_dir OUTPUT_DIR
                        output dir.

--am is acoustic model type with the format {model_name}_{dataset}
--am_config, --am_ckpt, --am_stat, --phones_dict --speaker_dict are arguments for acoustic model, which correspond to the 5 files in the fastspeech2 pretrained model.
--voc is vocoder type with the format {model_name}_{dataset}
--voc_config, --voc_ckpt, --voc_stat are arguments for vocoder, which correspond to the 3 files in the parallel wavegan pretrained model.
--lang is the model language, which can be zh or en.
--text is the text file, which contains sentences to synthesize.
--output_dir is the directory to save synthesized audio files.
--ngpu is the number of gpus to use, if ngpu == 0, use cpu.

Tips

If you want to get better audio quality, you can use more audios to finetune or change configuration parameters in conf/finetune.yaml.
More finetune results can be found on finetune-fastspeech2-for-csmsc.
The results show the effect on csmsc_mini: Freeze encoder > Non Frozen > Freeze encoder && duration_predictor.

14 KiB Raw Blame History