Binaural Speech Enhancement with Spatial Cue Preservation Utilising Simultaneous Masking

Abstract

Binaural multi-microphone noise reduction methods aim at noise suppression while preserving the spatial impression of the acoustic scene. Recently, a new binaural speech enhancement method was proposed which chooses per timefrequency (TF) tile either the enhanced target or a suppressed noisy version. The selection between the two is based on the input SNR per TF tile. In this paper we modify this method such that the selection mechanism is based on the output SNR. The proposed modification of deciding which TF tile is target-or
noise-dominated leads to choices, which are better aligned with simultaneous masking properties of the auditory system, and, hence, improves the performance over the initial version of the algorithm.