A comparative study of explicit frequency and conventional signal representations for speech recognition

Abstract

The performance of a speech recognition system depends to a large extent on the signal representation strategy. It is therefore important to evaluate various signal representations, with a view to comparing their relative performance on speech recognition tasks. Various comparative studies have been reported earlier in the literature in this respect. However, these studies are limited to comparing some subsets of representations on different kinds of recognition tasks. In this sense, they preclude a fair comparison of the representations on the same task. In contrast, we attempt here a systematic and fairly comprehensive comparison of signal representations along various dimensions (frequency and amplitude scales, speaker normalization, and two statistical classifiers) on a standard (TIMIT) speech database. This study shows that the line spectrum pair frequency representation augmented with spectral amplitudes yields the best recognition performance.