Estimation and Correction of Relative Transfer Function for Binaural Speech Separation Networks to Preserve Spatial Cues

Deep learning based approaches have achieved great success in mono-channel and multi-channel speech separation, but limited studies have focused on the binaural output, not even to mention the preservation of spatial cues. Existing speech separation networks preserve spatial cues by improving the signal-to-noise ratio (SNR) of the separated speech, regardless of the different requirements between reducing noise and preserving spatial cues. This work proposed a framework to optimize spatial cue preservation for binaural speech separation. It consisted of a relative transfer function (RTF) corrector that modified the distorted RTF of the separated speech into a correct one, and an RTF estimator to extract the correct RTF. A new RTF estimator was designed to obtain an accurate RTF. The framework was evaluated on a binaural version of WSJ0-2mix dataset, which was spatialized by anechoic head-related impulse responses. Experimental results showed that the proposed framework significantly reduced the interaural time difference (ITD) and interaural level difference (ILD) errors of the existing binaural separation networks, but did not notably sacrifice the SNR of the separated speech signals.