Anechoic Demixing under Phase Wraparound Conditions in Assisted Living Environments

Speech demixing algorithms are needed for Assisted Living (AL) environments, which are characterized by large inter-microphone spacings, and significant phase-wraparound. The DUET algorithm's low computational complexity and separation performance make it a good candidate source separation algorithm in AL scenarios. Phase wraparound due to large relative delays is a feature of AL microphone deployments. Large relative delays cause DUET to perform poorly when the microphones are located far apart. We present two extensions of DUET that improve demixing accuracy when relative delays are up to 200 samples, and the sampling frequency is F_s = 16 kHz. The two proposed algorithms are called Elevato-DUET and Elevato-AdRess. Numerical evaluations on real speech mixtures in AL-type environment demonstrate that under big delay scenarios, Elevato-DUET and Elevato-AdRess perform better relative delay estimation than the classical DUET. Our findings also highlight that Elevato-AdRess consistently surpasses Elevato-DUET across both subjective (BSS Eval, PESQ) and objective (SDR, SIR, SAR, MOS) demixing performance metrics, particularly in environments with inter-microphone gaps of up to 3.21 meters. However, it is worth noting that this performance improvement with Elevato-AdRess is accompanied by increased Matlab execution time.