Speech enhancement via Only Mostly Blind Source Separation

Abstract

In environments in which multiple simultaneously-active acoustic sources contribute to sensor responses, Blind Source Separation (BSS) signal processing techniques may be employed to separate (that is, estimate or reconstruct) the signal characteristics of "hidden" sources. Only Mostly Blind Source Separation (OMBSS) involves the estimation of similar sources in important contexts in which non-acoustic information is also available about one or more of the contributing sources. Recently-reported objective source separation performance measures confirm that non-acoustic information can be used effectively to support high-quality separation in situations in which traditional BSS methods perform poorly (e.g., when more sources are active than there are microphones available). Here we present the results of additional perceptual and objective tests showing that OMBSS processing enhances the intelligibility of speech recorded in the presence of multiple simultaneous speech-babble and non-speech maskers.