Results based on the M-Measure obtained for a single speech source with −30° DOA and a localized interferer (vacuum cleaner) with 40° DOA at varying SNR (−10 dB to 20 dB, abscissa). Top row: Results are shown for the office environment. Outcomes of stages B, E, and F in Figure 1. Left panel (A): DOA-dependent probability of sound source presence P _DOA(α) (Eq. (11)). Center panel (B): M-Measure (MM, Eq. (7)) obtained from phoneme posteriorgram output of a deep neural network for speech recognition as a measure of direction-dependent speech quality. Right panel (C): Combination of spatial source probability and ASR-based speech quality measure (Eq. (12)) recovers the speech source location from the mixture with the dominant interferer. Bottom row: Speech location detection performance (hit rate) based the information presented above and using Equations A: (14), B: (15), and C: (13) in the anechoic (blue bars) and the office (red bars) environment.