Phoneme-based continuous speech recognition without pre-segmentation

Abstract

Phonemes, modeled by 2-5 profiles sampled from parameter vector sequences, are used as basic recognition units in a continuous speech recognition system.The segmentation is achieved during the recognition process.Starting from several reliable islands, words decomposed into syllables are hypothesized and then located in a phoneme lattice.The phoneme recognition reliability is used to guide syllable location.In an application for single-speaker spoken Chinese recognition, using a 250-rule context-free grammar with 200 terminals, a 90% sentence and 99% phoneme recognition rate is obtained.