Intra-frame variability as a predictor of frame classifiability

Trond Skogstad,Torbjørn Svendsen

DOI: https://doi.org/10.21437/Interspeech.2010-491

Abstract:This paper examines the association between the variability of the speech signal inside an analysis frame and the relative diffi-culty of classifying that frame. We introduce a novel measure of speech frame variability and show through classification experiments that this measure is a strong predictor of classifiability, even when conditioning on the distance to segment boundaries. Finally, we show how to incorporate the measure as weights in the discriminant function of a GMM-HMM recognizer, thereby increasing the relative importance of low variability frames in both decoding and training. This is shown to give a reduction in error rates.