git clone https://github.com/Muhammad-Ahmad-Ghani/svoice_demo.git
cd svoice_demo
conda create -n svoice python=3.7 -y
conda activate svoice
# CUDA 11.3
conda install pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 cudatoolkit=11.3 -c pytorch -y
# CPU only
pip install torch==1.12.0+cpu torchvision==0.13.0+cpu torchaudio==0.12.0 --extra-index-url https://download.pytorch.org/whl/cpu
pip install -r requirements.txt

This is an intermediate checkpoint just for demo purpose.

create directory outputs/exp_ and save checkpoint there

svoice_demo
├── outputs
│   └── exp_
│       └── checkpoint.th
...

conda activate svoice
python demo.py

Create dataset mix_clean with sample rate 16K using librimix repo.

Dataset Structure

svoice_demo
├── Libri{NUM_OF_SPEAKERS}Mix_Dataset -> Libri7Mix_Dataset
│   └── wav{SAMPLE_RATE_VALUE}k -> wav16k
│       └── min
│       │   └── dev
│       │       └── ...
│       │   └── test
│       │       └── ...
│       │   └── train-360
│       │       └── ...
...

Run predefined scripts if you want.

# for 7 speakers
bash create_metadata_librimix7.sh
# for 10 speakers
bash create_metadata_librimix10.sh

Change conf/config.yaml according to your settings. Set C: NUM_OF_SPEAKERS value at line 66 for number of speakers.

python train.py

This will automaticlly read all the configurations from the conf/config.yaml file. To know more about the training you may refer to original svoice repo.

python train.py ddp=1

python -m svoice.evaluate <path to the model> <path to folder containing mix.json and all target separated channels json files s<ID>.json>

The svoice code is borrowed from original svoice repository. All rights of code are reserved by META Research.

@inproceedings{nachmani2020voice,
  title={Voice Separation with an Unknown Number of Multiple Speakers},
  author={Nachmani, Eliya and Adi, Yossi and Wolf, Lior},
  booktitle={Proceedings of the 37th international conference on Machine learning},
  year={2020}
}

@misc{cosentino2020librimix,
    title={LibriMix: An Open-Source Dataset for Generalizable Speech Separation},
    author={Joris Cosentino and Manuel Pariente and Samuele Cornell and Antoine Deleforge and Emmanuel Vincent},
    year={2020},
    eprint={2005.11262},
    archivePrefix={arXiv},
    primaryClass={eess.AS}
}

This repository is released under the CC-BY-NC-SA 4.0. license as found in the LICENSE file.

The file: svoice/models/sisnr_loss.py and svoice/data/preprocess.py were adapted from the kaituoxu/Conv-TasNet repository. It is an unofficial implementation of the Conv-TasNet: Surpassing Ideal Time-Frequency Magnitude Masking for Speech Separation paper, released under the MIT License. Additionally, several input manipulation functions were borrowed and modified from the yluo42/TAC repository, released under the CC BY-NC-SA 3.0 License.