A boilerplate for reproducible and transparent computer audition research that leverages nussl, a source separation library. This project provides boilerplate code for training neural network models to separate mixtures containing multiple speakers, music, and environmental sounds. It is easy to add and train new models or datasets (GPU sold separately). The goal of this project is to enable further reproducibility within the source separation community.
To start a new project:
cookiecutter gh:pseeth/cookiecutter-nussl
The documentation is here. It includes guides for getting started, training models, creating datasets, and API documentation.
The following models can be trained:
- Mask Inference
- Deep Clustering
- Chimera
- TasNet
on the following datasets:
- MUSDB18
- MIR-1k
- Slakh
- WSJ0-mix2
- Wham!
This project utilizes building block components from nussl for input/output
(reading/writing audio, STFT/iSTFT, masking, etc.), and for neural network construction
(recurrent networks, convolutional networks, etc) to train models with minimal setup.
The main source separation library, nussl, contains many pre-trained models trained
using this code. See the External File Zoo (EFZ)
for trained models.
This project uses
cookiecutter.
Cookiecutter is a logical, reasonably standardized, but flexible project structure
for doing and sharing research. This project and nussl are both built upon
the PyTorch machine learning framework, as such, building new
components is as simple as adding new PyTorch code, though writing python is not required.
- Install
cookiecuttercommand line:pip install cookiecutter(generates boilerplate code)
This project is licensed under the terms of the MIT License