Perceptual and computational separation of simultaneous vowels: Cues arising from low-frequency beating

Abstract

Identification of simultaneous speech sounds, such as pairs of steady-state vowels (double vowels), is more accurate when there is a difference in fundamental frequency (F0). Accuracy of identification for double vowels increases with increasing F0 difference (delta F0) asymptoting above 1 semitone. The experiment described here attempts to distinguish two mechanisms underlying this effect: first, perceptual separation by grouping together harmonic components of a common F0; and, second, exploitation of the fluctuations in the spectral envelope of the composite stimulus that result from beating between unresolved components. The beating is mainly caused by interactions between corresponding harmonics of the two vowels with a small delta F0. Identification accuracy for normal, harmonically excited double vowels was compared with that for double vowels composed from the same components, but whose constituent vowels were excited by a mixture of the two harmonic series. These double vowels were designed to produce similar beating patterns to the normal double vowels. Both harmonically and inharmonically excited constituents improved identification with increasing delta F0, but the increase was larger for harmonically excited vowels. A computational model based upon psychophysical measurements of auditory frequency and temporal resolution correctly predicted an increase in accuracy of identification with increasing delta F0 which was attributable to beating. The results are interpreted in terms of a spectral change cue in the identification of double vowels with delta F0's which complements grouping by F0, and which plays a dominant role for delta F0's smaller than 1 semitone.