iSTFTNet : Fast and Lightweight Mel-spectrogram Vocoder Incorporating Inverse Short-time Fourier Transform

This repository is based on the opensource implementation of iSTFTNet (model C8C8I). Our contribution to the repository:

• shared the weights of the model we trained on robust internal dataset consists of Russian speech recorded in different acoustic conditions with sample rate 22050 Hz;
• added loguru & wandb;
• added Docerfile for faster env set up;
• updated the code with several scripts to compute mel-spectrograms and convert the model to .onnx.

Note: according to our tests iSTFT Net shows even higher synthesis quality than HiFi GAN, with a 2x acceleration of RTF.

Table of Contents

• Setup env
• Inference
• Train
• ONNX
• Citations
• References

Setup env

Docker

bash run_docker.sh

Conda

conda create —name istft-vocoder python=3.10
pip install -r requirements.txt

Inference

Download checkpoints

bash download_checkpoints.sh

Your file structure should look like:

├── data                                                                                                                                                                                 
│   ├── awesome_checkpoints                                                                                                                                                              
│   │   ├── do_00975000                                                                                                                                                                  
│   │   ├── g_00975000                                                                                                                                                                   
│   │   └── g_00975000.onnx                                                                                                                                                              
│   ├── deep_voices_mel                                                                                                                                                                  
│   │   ├── andrey_preispolnilsya.npy                                                                                                                                                    
│   │   ├── egor_dora.npy
│   │   └── kirill_lunch.npy
│   └── deep_voices_wav
│       ├── andrey_preispolnilsya.wav
│       ├── egor_dora.wav
│       └── kirill_lunch.wav

Note: we trained the model with batch size 16 using 4 a100 GPUs for ~1M steps.

Filename	Description
do_00975000	Discriminator checkpoint.
g_00975000	Generator checkpoint.
g_00975000.onnx	.onnx model.
deep_voices_mel	Directory with 3 mel-spectrograms of test-audios.
deep_voices_wav	Directory with 3 original audios – voices of our team, this audios were not seen during the training.

Inference

To run inference with downloaded test-files:

python -m src.inference

To run inference with your own files or parameters:

Parameter	Description
config_path	Path to config.json.
input_wavs_dir	Directory with your wav files to synthesize, default is /app/data/deep_voices_wavs
input_mels_dir	Directory with pre-computed mel-spectrograms to synthesize mel. Note that mel-spectrograms should be computed with compute_mels_from_audio.py script, default is /app/data/deep_voices_mels.
compute_mels	Pass --no-compute_mels if you precomputed mels, if not specified mels will be computed from the audios in input_wavs_dir.
onnx_inference	If specified, checkpoint file should be .onnx file.
onnx_provider	Used if onnx_inference is specified, default provider is CPUExecutionProvider for CPU inference.
checkpoint_file	Path to the generator checkpoint or .onnx model.
output_dir	Path where generated wavs will be saved, default is /app/data/generated_files.

Train

To train the model:

1. Login from CLI to Wanb account: wandb login
2. Create train.txt and val.txt with create_manifests.py.
3. Run src.train

Parameters for training and finetuning the model:

Parameter	Description
input_training_file	Path to the train.txt.
input_validation_file	Path to the val.txt.
config_path	Path the config.json.
input_mels_dir	Path to the directory with mel-spectrograms, specify if you would like to train / finetune the model on Acoustic Model outputs.
fine_tuning	If specified will look for mel-spectrograms in input_mels_dir.
checkpoint_path	Path to the directory with checkpoints, if you would like to finetune the model on your data based on our checkpoints: /app/new_checkpoints.
training_epochs	N epochs to train the model.
wandb_log_interval	N steps through which log training loss to wandb.
checkpoint_interval	N steps through which save checkpoint.
log_audio_interval	N steps through which log generated audios from validation dataset to wandb.
validation_interval	N steps through which run validation and log validation loss to wandb.

Note: for correct inference and finetuning from our checkpoints, parameters: num_mels, n_fft, hop_size, win_size, sampling_rate, fmin and fmax should not be changed.

ONNX

Find the instructions to infer .onnx model in the Inference block. To convert trained model to .onnx:

python -m srcipts.convert_to_onnx

Parameter	Description
checkpoint_file	Path to the generator checkpoint.
config_path	Path to the config.json.
converted_model_path	Path where converted model will be saved, default is /app/istft_vocoder.onnx.

Citations

@inproceedings{kaneko2022istftnet,
title={{iSTFTNet}: Fast and Lightweight Mel-Spectrogram Vocoder Incorporating Inverse Short-Time Fourier Transform},
author={Takuhiro Kaneko and Kou Tanaka and Hirokazu Kameoka and Shogo Seki},
booktitle={ICASSP},
year={2022},
}

@misc{deepvk2023istft,
  author = {Daria, Diatlova},
  title = {istft-net},
  year = {2023},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {https://github.com/deepvk/istft-net}

References

• https://github.com/rishikksh20/iSTFTNet-pytorch