Pronunciation Modeling for Spontaneous Mandarin Speech Recognition

Pronunciation variations in spontaneous speech can be classified into complete changes and partial changes. A complete change is the replacement of a canonical phoneme by another alternative phone, such as ‘b’ being pronounced as ‘p’. Partial changes are variations within the phoneme such as nasalization, centralization and voiced. Most current work in pronunciation modeling for spontaneous Mandarin speech remains at the phone level and can model only complete changes, not partial changes. In this paper, we show that partial changes are much less clear-cut than previously assumed and cannot be modelled by mere representation by alternate phone units. We present a solution for modeling both complete changes and partial changes in spontaneous Mandarin speech. In order to model complete changes, we adapted the decision tree-based pronunciation modeling from English to Mandarin to predict alternate pronunciations. To solve the data sparseness problem, we used cross-domain data to estimate pronunciation variability. To discard the unreliable alternative pronunciations, we proposed a likelihood ratio test as a confidence measure to evaluate the degree of phonetic confusions. In order to model partial changes, we proposed partial change phone models (PCPM) with acoustic model reconstruction. PCPMs are regarded as extended units of standard phoneme or initial/final subword units, and can be used efficiently to represent partial changes. In order to avoid model confusion, we generated auxiliary decision trees for PCPM triphones, and used decision tree merge to perform acoustic model reconstruction. The effectiveness of these approaches was evaluated on the 1997 Hub4NE Mandarin Broadcast News corpus with different styles of speech. Our phone level pronunciation modeling provided an absolute 0.9% syllable error rate reduction, and the acoustic model reconstruction approach was more efficient than that to cover pronunciation variations, yielding a significant 2.39% absolute reduction in syllable error rate for spontaneous speech. In addition, our proposed method deals with partial changes at the acoustic model level and can be applied to any automatic speech recognition system based on subword units.