A PyTorch implementation of "Real-time Low-latency Music Source Separation using Hybrid Spectrogram-TasNet", published in ICASSP2024, by S. Venkatesh, A. Benilov, P. Coleman and F. Roskam (https://doi.org/10.48550/arXiv.2402.17701).
Mainly made for personal practice, some things might not be optimally implemented.
Create a dedicated conda environment then install the dependencies:
pip3 install -r requirements.txt
Some remarks:
- The spectrogram branch implemented here is magnitude-only, using the phase of the mixture to reconstruct complex STFTs. No particular reason behind this choice; might be better to stack real and imaginary parts.
- The model is trained using the open MUSDB18-HQ dataset (https://sigsep.github.io/datasets/musdb.html).
Use torchaudio.datasets.MUSDB_HQ to download the database, then run the script utils/preprocess_musdb.py to run preprocessing routines.