Nobutaka Ito's cSTMM unifies speech separation models, beats baseline by 0.25dB

In audio and speech processing, blind speech separation (BSS) aims to isolate individual speakers from a mixed recording without prior knowledge of the sources or the environment. Traditional mask-based BSS often relies on extracting phase and level difference features under specific assumptions like plane-wave propagation. Ito's work instead adopts a directional statistical approach that directly clusters normalized multichannel observations on the complex unit sphere.

The key innovation is cSTMM, a framework that unifies four previously disparate mixture models — the complex angular central Gaussian mixture model (cACGMM), complex Bingham mixture model (cBMM), complex Watson mixture model (cWMM) — through a single degrees-of-freedom parameter ν. This allows systematic exploration of how the tail behavior of the distribution affects separation quality. Ito derived a generalized minorization-maximization (MM) algorithm for parameter estimation. On noiseless, reverberated LibriSpeech data, the model achieved up to 0.25dB average SDRi improvement across all test conditions with ν=1, numerically recovering the other models at specific ν values.