A Comparative Study of Features for Continuous Speech Recognition in Real World Conditions

摘要

This paper is dedicated to the problem of selection of the best parameter set in a robust recognition experiment taking into account the environmental noise and the speaker task and emotional stress. The parameters are derived from LPC and cepstral analysis, in usual and perceptual versions. Two kinds of effectiveness comparisons were made: the LPC parameters versus the cepstral ones and the perceptual versus the usual parameters. In our tests, with a recognizer based on HMMs, the cepstral parameters perform in the most situations better versus the linear prediction ones, and the perceptual parameters better than the usual ones.