The modulation spectrum in the automatic recognition of speech
Hynek Heřmanský · 2002
The article questions the reliability of the short term spectral envelope as the dominant carrier of the phonetic identity of a given speech instant and suggests the temporal dynamics of components of the spectral envelopes as more reliable means for deriving the linguistic context of the speech message. It shows that analysis of the modulation spectrum offers a means for the systematic evaluation of medium term temporal dynamics of speech features. Such a medium term dynamic has been previously efficiently utilized in the computation of dynamic features and in RASTA processing. We aim for data driven analysis of the modulation spectrum and demonstrate the importance of syllabic rate modulation spectral components for speech communication.