AN AUDIO STEM EXTRACTOR

COMPUTER ENGINEERING

This project introduces Demucs (Deep Extractor for Music Sources), a cutting-edge music source separation model, particularly focusing on drums, bass, and vocals extraction from complex music mixtures. Demucs leverages a U-Net convolutional architecture inspired by Wave-U-Net, incorporating state-of-the-art techniques in deep learning. The latest version, Demucs v4, introduces the Hybrid Transformer Demucs, a novel approach employing a hybrid spectrogram/waveform separation model using Transformers. This model features a dual U-Net structure with a cross-domain Transformer, achieving a Signal-to-Distortion Ratio (SDR) of 9.00 dB on the MUSDB HQ test set. Additionally, by using sparse attention kernels and per-source fine-tuning, a state-of-the-art SDR of 9.20 dB is attained. The paper provides comprehensive insights into the architecture, training methodology, and performance evaluation of Demucs. It discusses model comparisons, system requirements, and practical instructions for using Demucs for music separation tasks. The release notes highlight significant updates and additions, including support for the SDX 2023 Challenge and integration with torchaudio. Moreover, the paper offers guidelines for training Demucs models and reproducing results from the MDX Challenge. To facilitate adoption, Demucs is made available as a Python package, compatible with various operating systems and environments. The paper concludes with citations for proper attribution and licensing information. Demucs represents a significant advancement in music source separation, offering researchers and practitioners a powerful tool for audio processing and analysis tasks.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.