Real-time Low-latency Music Source Separation using Hybrid Spectrogram-TasNet
Satvik Venkatesh, Arthur Benilov, Philip Coleman, Frederic Roskam
TL;DR
This work tackles real-time, low-latency music source separation by adapting existing MSS models and introducing a novel hybrid architecture, HS-TasNet, that leverages both spectral and waveform information. By enforcing causal, low-latency constraints (23 ms), the authors demonstrate that HS-TasNet achieves an SDR of 4.65 on MusDB, rising to 5.55 with additional training data, and outperforms a real-time TasNet under the same conditions. They also present a lighter HS-TasNet-Small variant that reduces parameters and maintains strong performance, with favorable inference times on CPU. The findings suggest that efficient, hybrid approaches can enable practical real-time demixing for applications like hearing aids, remixing, and live audio streams, while offering directions for further latency reductions and training strategies.
Abstract
There have been significant advances in deep learning for music demixing in recent years. However, there has been little attention given to how these neural networks can be adapted for real-time low-latency applications, which could be helpful for hearing aids, remixing audio streams and live shows. In this paper, we investigate the various challenges involved in adapting current demixing models in the literature for this use case. Subsequently, inspired by the Hybrid Demucs architecture, we propose the Hybrid Spectrogram Time-domain Audio Separation Network HS-TasNet, which utilises the advantages of spectral and waveform domains. For a latency of 23 ms, the HS-TasNet obtains an overall signal-to-distortion ratio (SDR) of 4.65 on the MusDB test set, and increases to 5.55 with additional training data. These results demonstrate the potential of efficient demixing for real-time low-latency music applications.