Plosive Discrimination by Running Spectra in 40-ms Initial Segment.

Abstract

The present paper deals with invariant features for place of articulation. This paper describes results of consonant discrimination using 5 running spectra extracted during burst after 40-ms waveform of Japanese plosive-vowel syllables. The number of talkers analyzed was 48 and the phonetic environments were /p, t, k, b, d, g/ and five Japanese vowels. The features were extracted so as to be invariant with respect to both talker and vowel : First, a set of canonical variables was selected and transformed from spectrum variables in each frame, and then these canonical variables were integrated and used for final discrimination. Due to integration of consonantal features in these successive spectra, correct classification score was higher than that of a single spectrum experiment. We examined the feature distribution in the burst and the following transition spectrum, and the differences between voiced and voiceless plosives.