Naxi-accented Mandarin speech recognition based on pronunciation dictionary adaptation

This paper primarily concentrates on Mandarin LVCSR (Large Vocabulary Continuous Speech Recognition) for nonnative speakers from Yunnan, China. With an available standard Mandarin speech recognizer, we attempt to establish an accent-specific speech recognizer for the speakers whose native language is Naxi language, based on the Initial-Final structure of the Chinese language, in combination with the variation regularity of pronunciation in Naxi-accented speech. In order to obtain the variation regularity of the initials, finals and syllables in Naxi-accented speech, we analyze the pronunciation variability of the non-native speech using the data-driven approach supervised by expert knowledge. A novel method to automatically construct multi-pronunciation dictionary of the given accent which can be easily extended to the other linguistic minorities´ accents was proposed. Experimental results show that the Naxi-accented speech recognition rates with the multi-pronunciation dictionary constructed by the proposed method are higher than the rates with single pronunciation dictionary, after imported bi-gram language model.