Illustration of the proposed system: (A, Sect. 2.1) A virtual acoustic scene (speaker at −30° and either a spatially diffuse noise or localized noise source at 40°) is captured with a 6-channel behind-the-ear hearing aid. (B, Sect. 2.2) Four channels are used to extract features for sound source localization fed to an SVM-based classifier. This results in a probability for direction of arrival of localized sound sources in the acoustic scene over time, which are averaged over time. (C, Sect. 2.3) A six-channel beamformer is used for spatial signal enhancement. (D, Sect. 2.4) ASR features are extracted from the beamformer signals and used as input to a DNN trained on speech data. This results in phoneme probabilities over time for each beamforming direction. (E, Sect. 2.5) An ASR-based speech quality measure is applied to these probabilities. (F, Sect. 2.6) By integrating information from two processing streams, the speaker at −30° (solid line) is clearly separated from the localized noise (dashed line).