Vocal Isolation with SpeechBrain

This project explores music source separation for vocal isolation using Conv-TasNet and SepFormer built with the SpeechBrain toolkit.

Project Overview

The repository contains a notebook-driven course project and a refactored training layout for publishing on GitHub. The workflow covers:

• MUSDB18 data preparation
• metadata CSV creation
• model training with SpeechBrain
• evaluation and qualitative listening

Repository Structure

.
├── configs/
│   ├── convtasnet.yaml
│   ├── sepformer.yaml
│   └── *_original.yaml
├── notebooks/
│   └── Project.ipynb
├── scripts/
│   └── prepare_musdb.py
├── src/
│   ├── train.py
│   └── train_original.py
├── data/
│   ├── raw/
│   └── processed/
├── results/
├── requirements.txt
└── README.md

Installation

pip install -r requirements.txt

Data Preparation

1. Download the MUSDB18 dataset.
2. Place the raw dataset under data/raw/.
3. Generate musdb_train.csv, musdb_valid.csv, and musdb_test.csv from the preprocessing logic in notebooks/Project.ipynb.

Starter command:

python scripts/prepare_musdb.py

Training

Train Conv-TasNet:

python src/train.py configs/convtasnet.yaml

Train SepFormer:

python src/train.py configs/sepformer.yaml

Notes

• The original uploaded files are preserved as *_original.yaml and train_original.py.
• The refactored files are intended to make the repository easier to read and maintain.
• You may still need to adapt dataset loading hooks depending on how your CSV manifests are generated.

Acknowledgment

This project builds on the SpeechBrain toolkit for speech and audio processing.

SpeechBrain repository:

• SpeechBrain GitHub: https://github.com/speechbrain/speechbrain

Please cite SpeechBrain if you use this toolkit:

@article{speechbrain2021,
  title={SpeechBrain: A General-Purpose Speech Toolkit},
  author={Ravanelli, Mirco and others},
  journal={arXiv preprint arXiv:2106.04624},
  year={2021}
}