Component structuring and trajectory modeling for speech recognition

Abstract

When the speech data are produced by speakers of different age and gender, the acoustic variability of any given phonetic unit becomes large, which degrades speech recognition performance. A way to go beyond the conventional Hidden Markov Model is to explicitly include speaker class information in the modeling. Speaker classes can be obtained by unsupervised clustering of the speech utterances.This paper introduces a structuring of the Gaussian components of the GMM densities with respect to speaker classes. In a first approach, the structuring of the Gaussian components is combined with speaker class-dependent mixture weights. In a second approach, the structuring is used with mixture transition matrices, which add dependencies between Gaussian components of mixture densities (as in stranded GMMs). The different approaches are evaluated and compared in detail on the TIDIG-ITS task. Significant improvements are obtained using the proposed approaches based on structured components. Additional results are reported for phonetic decoding on the NEOLOGOS database, a large corpus of French telephone data.