On Multi-input Multi-frame MVDR Filter for Speech Enhancement with Heterophasic Presentation

Multi-channel speech enhancement attempts to recover a target speech signal from noisy observations by exploiting spatial information captured by a microphone array. Conventional approaches typically produce a single output that contains both the desired speech and some residual noise, which neglects the benefits of human binaural hearing system. To overcome this limitation, we propose in this work a multi-input multi-frame binaural-output (MIMFBO) noise reduction method operating in the short-time-Fourier-transform (STFT) domain. This method utilizes both inter-channel and inter-frame correlations to design binaural filters that maximize the interaural coherence (IC) of the desired speech signal while minimizing the IC of the noise, all under distortionless constraints for the desired target speech. As a result, the perceived target signal and residual noise are spatially separated, substantially enhancing speech intelligibility. Simulation results demonstrate the proposed method’s superiority, showing significant improvements in PESQ scores over both the single-input binaural-output MVDR and multi-input binaural-output MVDR approaches. Moreover, subjective listening tests confirm its perceptual benefits.