SBSS: A research toolkit for scalable blind source separation

We present SBSS, a research toolkit for scalable blind source separation (BSS). Statistical BSS has been investigated to extract source signals from a multichannel mixture with minimum prior information. A typical approach to BSS is formulating a probabilistic generative model of a multichannel mixture and estimating its parameters through statistical inference. Recent approaches, such as reverberation as supervision (RAS) and neural full-rank spatial covariance analysis (neural FCA), have combined neural networks with BSS to enable training from real audio recordings, where isolated source signals are difficult to obtain. Despite their effectiveness, implementing these methods remains challenging due to the need for expertise in both neural and statistical signal processing. To accelerate research in neural BSS, we developed SBSS, which provides modular components and training recipes for state-of-the-art methods. It includes differentiable implementations of BSS algorithms and reproducible recipes for existing studies. We demonstrate that SBSS serves as a flexible platform to extend recent BSS methods for better performance.