Distributed FastMNMF slashes blind source separation compute by using block-diagonal covariances

Blind source separation (BSS) with distributed microphone arrays — such as arrays in smart rooms or sensor networks — typically requires either heavy computation (processing all microphones jointly) or poor performance (processing only one subarray). A new paper from Hirotaka Nishikori and colleagues at the University of Tokyo introduces distributed FastMNMF, a method that strikes a practical balance. The key innovation is enforcing a block-diagonal structure on the source spatial covariance matrices, which lets matrix inversions happen independently within each subarray. Meanwhile, the NMF-based source spectrogram model is shared across the entire array, aggregating activity information from all microphones while discarding inter-subarray covariance details.

In synchronized, noiseless simulations with fixed room and source geometry, distributed FastMNMF required less computation time than conventional full-array FastMNMF and achieved higher average source-to-distortion ratios than single-subarray FastMNMF. Critically, it worked in a five-source scenario where each four-microphone subarray was locally underdetermined — something a single-subarray approach cannot handle. The trade-off is that it ignores phase relationships between subarrays, but the results suggest this cost is acceptable for many practical applications. The paper is available on arXiv (2605.19388) and has implications for smart assistants, teleconferencing, and acoustic monitoring over large areas.