This dissertation describes automatic speech recognition (ASR) systems for continuous speech of Mandarin Chinese which are based upon the technologies of spectral/temporal features, tone modeling techniques and their integration with hidden Markov model based recognizers.;Standard DCTC/DCSC features are first optimized by advanced techniques for the purpose of improving the overall recognition rate. The pitch feature and the general spectral energy distribution and their correlations to tones are also discussed.;Since the training data for this research is relatively limited (320 minutes of continuous speech from 20 speakers), the recognition system is based on a phonetic recognizer. Two major strategies are used for phonetic recognition for Mandarin speech: explicit tone modeling (e.g. tones are modeled separately from base syllables by HMMs) and implicit tone modeling (e.g. tones are considered as a part of syllable FINALS and modeled together). Finally different strategies for synchronizing the tone recognizer and base syllable recognizer are investigated and evaluated at the tonal syllabic level in this dissertation.
展开▼
