On the semantic content of intelligibility test material on assessing signal processing algorithms

Abstract

Signal processing algorithms can be used to enhance speech intelligibility by improving the detectability of important acoustic features. It should follow from this that other nonacoustic portions of the signal are a source of experimental error that should be controlled. Nevertheless, certain signal processing techniques produce reliable improvements in intelligibility only when that speech is meaningful. A series of intelligibility tests was carried out with co‐channel speech, which combines the speech of two talkers into a single channel. This speech had been processed using the harmonic magnitude suppression (HMS) technique [Naylor and Boll, ICASSP (1987)] to diminish the effect of the louder talker. Intelligibility testing using target speech of read PB sentences masking read four‐word nonsense sentences revealed no differences between treated and untreated speech. Intelligibility using complete sentences as both masker and target showed significant differences in favor of the HMS treated speech. An analysis of signals and errors will be presented to show how listeners are integrating information from nonsignal sources to improve performance.