Objective measures of blind source separation and the intelligibility of separated speech

Abstract

When multiple acoustic sources are present in an environment monitored by multiple microphones, each microphone’s response signal is typically a weighted and delayed mixture of the active source signals. Subject to important constraints, these mixtures can be processed using Blind Source Separation (BSS) methods to construct output signals representative of each of the “hidden” acoustic sources in isolation. An important objective measure of the quality of a given “separation solution” is the amount of mutual information remaining in the BSS output channels: the less mutual information across channels, the better the separation. Often, the acoustic signals of interest in a complex acoustic environment are speech sources, and the measure of separation quality that really matters to listeners is speech intelligibility—a subjective (perceptual) measure that cannot be directly computed by a signal processing system. Thus, it would be useful to know how well objective measures of separation quality, such as mutual information, correlate with intelligibility. We present results of perceptual and objective tests exploring the relationship between objective separation metrics and word intelligibility as estimated by listeners’ responses in word identification tasks, for target-word-in-carrier-phrase utterances recorded in an acoustic environment with masking speech babble and broadband noise sources.