Modelling speech in the wavelet domain

MILONE, D. H.

Rosario

Workshop; Workshop on Signal Processing; 2006

IEEE

In the last decade wavelet transform has shown its potential as representation of signals and images for a wide range of applications. Nowadays, besides it continues being an active research field, some of these applications have given rise important technological developments. The discrete wavelet transform has been applied to traditional problems in speech processing, such as enhancement, compression, detection, recognition, synthesis, pitch tracking and speaker identification. However, very different levels of success have been obtained for each of those applications, mainly since there are still some basics limitations to obtain an appropriated model of speech in the wavelet domain. Within them, we can mention the inadequate frequency resolution in those bands where important cues of speech are coded. Other problems are related with the short term analysis widely used in speech technologies. In this context, we can mention the lack of a suitable statistical model for speech in a wavelet by-frame analysis and the temporal shifts that, at each scale, are generated by the relative phase between the frame window and the fundamental frequency and harmonics of speech.