Robust Speech Features based on synchrony spectrum determination using PLLs

Abstract

In this work we propose to include synchrony effects, known to exist in the auditory system, to represent speech signal information in a robust way. The system decomposes the signal in a number of simpler signals, and utilizes a bank of Phase Locked Loops (PLLs) to obtain information of the frequencies present at each time. This information is interpolated in order to obtain a spectral-like representation based in synchrony effects, measured by the PLLs. Noisy speech recognition experiments are performed using this synchrony-based spectrum, which is transformed into a small set of coefficients by using a similar transformation as the one utilized for the Mel cepstrum features. We show their recognition performance compared to Mel cepstrum features obtained from the standard power spectrum. Some recognition improvements are obtained for the case of vocalic sounds for this approach, especially in the case of severe noise conditions.