Novel binaural cue loss preserves spatial hearing in DNN speech enhancement

Binaural speech enhancement for hearing aids reduces noise while preserving interaural cues—ILD and IPD—that are critical for spatial localization. While deep neural networks (DNNs) achieve excellent noise reduction, they often distort the relationship between left and right signals, hurting spatial awareness. In this paper, Jayteerth Amble and colleagues propose two novel binaural cue preservation losses to fix this distortion.

The first is a binaural reconstruction error loss that directly penalizes masking-induced distortion between left and right spectra, providing a more direct measure of binaural consistency than conventional separate ILD and IPD errors. The second is a binaural cue loss that jointly models ILD and IPD to better preserve binaural structure. Experimental results show that both losses maintain strong noise reduction performance while reducing masking-induced distortion compared to the state-of-the-art baseline. The second joint loss also outperforms the baseline in ILD preservation, making it a promising step for practical hearing aid algorithms.