Optimum Soft Mask for Monaural Speech Separation System

摘要

Monaural speech separation aims to separate the target speech from speech mixture recorded by single mike. The ideal binary mask (IBM) has been projected as a procedure goal in computational auditory scene analysis (CASA) based monaural speech separation. The IBM is essentially a matrix of binary numbers, the binary value 1 is allotted to the mask if the native signal-to-noise ratio (SNR) of a specific time-frequency (T-F) units exceeds the native criterion (LC) otherwise the value 0 is allotted to the mask. The IBM based speech synthesis could discard some components of the speech and leads to associate unnatural sound known as musical noise. This analysis work proposes the optimum soft mask (OSM) as an alternate to IBM to cut back the musical noise, by commutation the arduous limiting weights (i.e. 1 or 0) with the variable weights between 0 and 1. The IEEE speech corpus and NOISEX92 noises are wont to appraise the performance of projected optimum soft mask in terms of signal-to-noise ratio (SNR) and also the sensory activity analysis of speech quality (PESQ). The experimental results indicate the superior performance of the projected optimum soft mask as compared to the IBM and ideal multi-threshold mask (IMM) within the context of monaural speech separation.