When the speech data is produced by speakers of different age and gender, the acoustic variability of any given phonetic unit becomes large, which degrades speech recognition performance. One way to go beyond conventional Hidden Markov Model is to explicitly include speaker class information in the modeling. Speaker classes can be obtained automatically, and they are used for building speaker class-specific acoustic models. This paper introduces a structuring of the Gaussian components of the GMM densities with respect to the speaker classes. In a first approach, this structuring of the Gaussian components is completed with speaker class-dependent mixture weights, and in a second approach, with transition matrices, which add dependencies between Gaussian components of mixture densities (as in stranded GMMs). The two approaches bring substantial performance improvements when recognizing adult and child speech. Using class-structured components plus mixture transition matrices reduces by more than one third the word error rate on the TIDIGIT corpus.

EXPLOR_STEP=$WICRI_ROOT/Wicri/Lorraine/explor/InforLorV4/Data/Main/Merge

HfdSelect -h $EXPLOR_STEP/biblio.hfd -nk 000A90 | SxmlIndent | more

HfdSelect -h $EXPLOR_AREA/Data/Main/Merge/biblio.hfd -nk 000A90 | SxmlIndent | more

{{Explor lien
   |wiki=    Wicri/Lorraine
   |area=    InforLorV4
   |flux=    Main
   |étape=   Merge
   |type=    RBID
   |clé=     Hal:hal-01080343
   |texte=   Explicit trajectories and speaker class modeling for child and adult speech recognition
}}