Discrete Token Modeling for Multi-Stem Music Source Separation with Language Models

This project represents a shift in Music Source Separation (MSS) from traditional signal-masking approaches (like Demucs or MDX-Net) to generative modeling using discrete tokens. While the methodology is technically sophisticated—utilizing a dual-path neural audio codec (HCodec) and a decoder-only LM—the defensibility is currently low (score 3) because it is a fresh research release with no established user base or community moat (0 stars, 7 forks). The frontier risk is high because major players like Meta (creators of Demucs) and ByteDance are already heavily invested in MSS; moving to a generative/token-based approach is a logical architectural evolution that these labs can replicate or improve upon quickly. The primary innovation is the application of the 'Audio Language Model' paradigm to separation, which helps with signal reconstruction in complex overlaps but introduces significant inference latency compared to standard U-Net/Conformer models. Commercial viability depends on whether the generative approach significantly outperforms existing SOTA models like Demucs v4 on benchmark SDR (Signal-to-Distortion Ratio) without introducing hallucinations, a common pitfall of AR models in audio separation.