IGA Speech Enhancement with Variance-Reduced Training

Abstract:
In our previous publication, we proposed a data-driven speech enhancement with so-called ideal gain averaging (IGA) weighting rules to estimate the clean speech spectra. Being implemented as a table look-up, the subbandindividual weighting rules were trained separately for speech presence and speech absence by taking the average of all ideal gains computed from clean speech and noise training signals recorded in the environment of interest. In this contribution we present a new training methodology selecting appropriate ideal gains to compute the final IGA weighting rules for speech presence and speech absence. This selection of ideal gains effectively reduces the bias of the weighting rules under mediumand low SNR conditions, which occurs due to the imperfect voice activity detection (VAD) computation. Compared to our previous publication, the proposed training methodology yields an improvement in terms of speech preservation and noise attenuation.