Skip to main content

Research Repository

Advanced Search

Linguistic constraints on speech recognition

Linguistic constraints on speech recognition Thumbnail


Abstract

This thesis is a study of the influence of linguistic constraints on automatic speech recognition by computer. The strategy is to constrain the recognition process by knowledge about the pattern to be recognised, in this case the intonation system of British English.
The categorical nature of the perception of pitch movement corresponding to nuclear syllable intonation is demonstrated. It is shown that Halliday's system of five primary tones is appropriate and applicable to automatic intonation analysis.
A computer analysis system was constructed which uses dynamic programming time warping to compare fundamental frequency patterns. The analysis is constrained by an intonation tone group structure grammar. The grammar consists of context-free rewrite rules and a lexicon of intonation templates. The analysis system comprises a rule translator, a syntax-directed analyser, dynamic programming fundamental frequency contour matcher, and a speech preprocessor.
The system was used in nuclear tone analysis and classification experiments for speaker dependent and independent tone recognition, and for connected utterance analysis over the complete tone group.
The results show that a limited prosodics-only speech recogniser is practical.

Publicly Available Date Mar 29, 2024

Files




Downloadable Citations