Stemuc Audio Forge: AI-based Music Source Separation Using Demucs and CUDA Acceleration

Raphael Serraino Theil Meres, Thiago Silva de Souza, Rigel Procópio Fernandes, Cassius M. do C. Figueiredo

DOI: 10.14209/sbrt.2025.1571157327

Evento: XLIII Simpósio Brasileiro de Telecomunicações e Processamento de Sinais (SBrT2025)

Keywords:

Abstract

Audio source separation is a fundamental task in music information retrieval and is widely employed by musicians and audio engineers. This paper introduces Stemuc Audio Forge, a system that leverages the Demucs neural network to separate music into distinct stems (vocals, drums, bass, guitar, piano, and others). The system incorporates graphics processing unit (GPU) acceleration via CUDA, reducing the processing time from approximately five minutes on a CPU to less than 10 seconds on a GPU. Evaluation on the MUSDB18 dataset demonstrates high-quality stem separation and significant performance gains, making advanced music source separation feasible for real-world applications.

Download