A simple correlation-based model of intelligibility for nonlinear speech enhancement and separation

Applying a binary mask to a pure noise signal can result in speech that is highly intelligible, despite the absence of any of the target speech signal. Therefore, to estimate the intelligibility benefit of highly nonlinear speech enhancement techniques, we contend that SNR is not useful; instead we propose a measure based on the similarity between the time-varying spectral envelopes of target speech and system output, as measured by correlation. As with previous correlation-based intelligibility measures, our system can broadly match subjective intelligibility for a range of enhanced signals. Our system, however, is notably simpler and we explain the practical motivation behind each stage. This measure, freely available as a small Matlab implementation, can provide a more meaningful evaluation measure for nonlinear speech enhancement systems, as well as providing a transparent objective function for the optimization of such systems.