Discriminative weighting of multi-resolution sub-band cepstral features for speech recognition

Abstract

This paper explores possible strategies for the recombination of independent multi-resolution sub-band based recognisers. The multi-resolution approach is based on the premise that additional cues for phonetic discrimination may exist in the spectral correlates of a particular sub-band, but not in another. Weights are derived via discriminative training using the ‘Minimum Classification Error’ (MCE) criterion on loglikelihood scores. Using this criterion the weights for correct and competing classes are adjusted in opposite directions, thus conveying the sense of enforcing separation of confusable classes. Discriminative re-combination is shown to provide significant increases for both phone classification and continuous recognition tasks on the TIMIT database. Weighted recombination of independent multi-resolution subband models is also shown to provide robustness improvements in broadband noise.