English Alphabet Recognition Based on Chinese Acoustic Modeling

How to effectively recognize English letters spoken by Chinese people is our major concern in the paper. Some efforts are made to build Chinese extended Initial/Final (XIF) based HMMs for English alphabet recognition which can be integrated with large vocabulary continuous Chinese speech recognition (Chinese LVCSR) system based on a same XIF set. The alphabet-specific XIF HMMs are built using context-dependent modeling, decision tree based state clustering method, state-based phonetic mixture tying and pronunciation modeling techniques. Experiments have been done over a 32-speaker test set. Compared with English phoneme-based acoustic modeling, our proposed method can achieve a relative letter error rate reduction of 5.3% with a letter correctness of 97.2% for Chinese-accented English alphabet recognition. What’s more, the XIF-based HMMs for English alphabet can be integrated with Chinese LVCSR seamlessly to recognize Chinese as well as English letters simultaneously.