MSST: Open-source framework unifies music stem separation with LoRA and sliding-window inference

Music Source Separation (MSS) — isolating vocals, drums, bass, and other stems from a mixed track — is crucial for karaoke, remixing, and audio restoration. The quality depends on many engineering decisions: model choice, data preparation, loss functions, training configs, and post-processing. The new MSST framework, introduced by Roman Solovyev and colleagues, unifies training, validation, and inference for modern demixing models. It is fully open-source and configured through YAML files, supporting architectures like Demucs, Spleeter, and others. The framework includes practical enhancements such as sliding-window inference with cross-fading to smooth overlapping segments, test-time augmentation (TTA) to improve robustness, model ensembling for higher accuracy, and Low-Rank Adaptation (LoRA) for efficient fine-tuning.

MSST's ablation studies show measurable improvements from these techniques, making state-of-the-art separation more accessible. By consolidating the entire pipeline into a reproducible, configuration-driven interface, the framework enables rapid iteration and systematic experimentation. This is a significant step for both researchers and practitioners: it lowers the barrier to entry for building custom demixing models and allows easy comparison across different approaches. The code is available on arXiv (arXiv:2607.23395) and is expected to become a standard tool in the audio AI community.