A Single-Input/Binaural-Output Perceptual Rendering Based Speech Separation Method in Noisy Environments

In this paper, we address the challenge of single-channel speech separation in noisy environments, where two active speakers and background noise are present in the observed signal. We propose using a dual path recursive neural network (DPRNN) to estimate the desired binaural signals from the single-channel noisy input. When the estimated binaural signal is played through headsets, listeners perceive the two speakers as originating from opposite directions, with the background noise coming from a separate direction. Additionally, the background noise is perceived to be further away from the two speakers, resulting in an improved signal-to-noise ratio (SNR). Research in psychoacoustics indicates that spatial unmasking in the perceptual domain enhances speech intelligibility in complex auditory scenes. This hypothesis is supported by both subjective and objective evaluations, including a significant 26% improvement in modified rhyme test (MRT) scores reported in this paper.