Pronunciation Recognition and Assessment for Mandarin Chinese

This paper establishes a speaker-independent pronunciation recognition and assessment system with 673 words for mandarin Chinese under the background of a Chinese learning system framework. The recognition part is based on HTK using HMM (Hidden Markov Models) and improved in the aspect of acoustic model. Making use of the recognition results and the log-likelihood obtained from the Viterbi coding, the machine scoring is performed to evaluate the quality of pronunciation. Finally the assessment approach based on log probability is combined with the pronunciation recognition system established. The machine score and its correlation with expert score are discussed and the experimental result shows that the scores are close to people´s subjective scoring.