Comparison of Parametric Spectral Representations for Voice Recognition in Noisy Environments

Abstract

Voice recognition systems provide good performances when the speech signal is recorded in good conditions: low noise level, good microphones. But results are not sufficient for several real life noisy situations (e.g cars). The aim of the presented work was to compare three techniques of spectral parametrisation in terms of performances for speech recognition and more precisely to evaluate their robustness in noise. This study was part of a French GRECO project on the comparison of methods of parametric and non parametric spectral analysis for speech recognition. This project has used the existing speech recognition program SAMREC-1 with the speech data base EUROMO and the RSG_10 noise data base. The different techniques are evaluated by their scores of recognition in lexical accesses, for speaker dependent isolated word recognition based on Dynamic Time Warping.