MAP-based cross-language adaptation augmented by linguistic knowledge: from English to Chinese

Construction of a recognizer in a new target language usually involves collection of a comprehensive database in that language as well as manual annotation and model training. For rapid development of new language recognizer, we propose (1) substituting target language phoneme models by source language phoneme models trained previously; and (2) adapting target language phoneme models from source language phoneme models using maximum a posteriori method. Evaluations show that adaptation method results in an error reduction rate of 51% from using substitution, and an error reduction rate of 10% from language-dependent training. We also propose using data-driven method with linguistic knowledge, instead of heuristic rules, to align source language phonemes to target language phonemes, both for substitution and adaptation. Evalution shows that using linguistic rules brings a 2% increase in recognition rate.