vlozg / voicefilter

Unofficial PyTorch implementation of Google AI's VoiceFilter system

VoiceFilter

Dependencies

1. Python and packages

   This code was tested on Python 3.8 with PyTorch 1.10.0. Other packages can be installed by:

   pip install -r requirements.txt

Prepare Dataset

1. Download LibriSpeech dataset

   Install axel first (apt install axel).

   Use axel to download datasets.

   axel -n 10 -a -c "https://www.openslr.org/resources/12/train-clean-100.tar.gz"
   axel -n 10 -a -c "https://www.openslr.org/resources/12/train-clean-360.tar.gz"
   axel -n 10 -a -c "https://www.openslr.org/resources/12/dev-clean.tar.gz"
   axel -n 10 -a -c "https://www.openslr.org/resources/12/test-clean.tar.gz"

   Then, unzip tar.gz file to datasets folder:

   tar -xvzf train-clear-100.tar.gz
   tar -xvzf train-clear-360.tar.gz
   tar -xvzf dev-clean.tar.gz
   tar -xvzf test-clean.tar.gz

2. Edit config.yaml

   cd config
   cp default.yaml config.yaml

Train VoiceFilter

1. Get pretrained model for speaker recognition system

   The model can be downloaded at this GDrive link.

   Using gdown command for convenient download (gdown was installed via pip).

   gdown --id 1YFmhmUok-W76JkrfA0fzQt3c-ZsfiwfL

2. Run

   After specifying train_dir, test_dir at config.yaml, run:

   python train.py -c [config.yaml] -e [path of embedder pt file] -m [name] --train_set [list of datasets used to generate train data] --test_set [list of datasets used to generate test data]

   This will create chkpt/name and logs/name at base directory(-b option, . in default)

   For reproducing the original experiment, use: the following bash command:

   python train.py -c config.yaml -e embedder.pt -m powlaw_loss --train_set librispeech-train --test_set librispeech-test

   Supported dataset include (for detail implementation, see the source code in datasets/GenerateDataset.py): librispeech-train, librispeech-test, vctk , vin, voxceleb1-train, voxceleb1-test, voxceleb2-train , voxceleb1-test, zalo-train, zalo-test

3. View tensorboardX

   tensorboard --logdir ./logs

4. Resuming from checkpoint

   python trainer.py -c [config yaml] --checkpoint_path [chkpt/name/chkpt_{step}.pt] -e [path of embedder pt file] -m name --train_set [list of datasets used to generate train data] --test_set [list of datasets used to generate test data]

   For example, finetune with VN dataset:

   python train.py -c config.yaml -e embedder.pt -m powlaw_loss_finetune --checkpoint_path chkpt/powlaw_loss/chkpt_168000.pt --train_set vin zalo-train --test_set zalo-test

License

Apache License 2.0

This repository contains codes adapted/copied from the followings:

• utils/adabound.py from https://github.com/Luolc/AdaBound (Apache License 2.0)
• utils/audio.py from https://github.com/keithito/tacotron (MIT License)
• utils/hparams.py from https://github.com/HarryVolek/PyTorch_Speaker_Verification (No License specified)
• utils/normalize-resample.sh from https://unix.stackexchange.com/a/216475