Multiresolution Information Measures applied to Speech Recognition

TORRES, M. E.; RUFINER, H. L.; MILONE, D. H.; CHERNIZ, A.

Mendoza

Simposio; 7th Argentine Symposium on Computing Technology; 2006

SADIO

Considerable advances in automatic speech recognition have been made through application of hidden Markov models and development of different speech signal parametrization techniques. However, deterioration in the speech recognizers performance have been observed when systems trained with clean signals are tested with noisy signals. In this paper we present the extension of the continuous multiresolution entropy to different divergences and we introduce these information measures as new dimensions to the front–end stage of an ASR system. These new parameters take into account information about the changes in the dynamics of speech signal for different scales and is concatenated to a MFCC classic parametrization. Proposed methods are tested with speech signals corrupted with babble and white noise and compared with a classical mel cepstra parametrization. Results suggest that multiresolution information measures provide valuable information to the speech recognition system and could be included as an extra component in the pre-processing stage.