Combination of Acoustic Models Trained from Different Unit Sets for Chinese Continuous Speech Recognition

Combination of acoustic models trained from different unit sets is studied in this paper. For Chinese continuous speech recognition, Prevailing unit sets include context-dependent initial-final unit set and context-dependent phone unit set. Through experiments it is discovered that some Chinese syllables have higher recognition rates under initial-final model while some have higher recognition rates under phone model. In this paper, a method is proposed to combine these two acoustic models. On one hand the two acoustic models can be fully utilized during the recognition process; on the other hand, some models that lead to low recognition rate will not be used. Experiments show that in comparison with initial-final model and phone model, syllable error rate is reduced by 9.60% and 6.10% respectively after using the provided method.