Research On Chinese Continuous Speech Tone And Digit String Recognition System

Speech recognition has important application prospect in human-computer interaction, the communication, the internet and industrial control. Tone is important part of Chinese syllables, which plays word formation and meaning recognition role. In Chinese speech recognition technology, speak independent and continuous speech is research focus and difficulty at present. With the development of speech recognition technology, the tone recognition research has become one of the breakthrough directions.First, tone extraction algorithm has been researched. In this paper, multi-parameters (short-time energy, short-time zero-crossing rate and autocorrelation function values) combination has been adopted to judge the unvoiced/voiced of speech signal. Then circular average magnitude difference function has been used to calculate pitch. After extracting seven-dimensional feature parameters from tone curve, Chinese continuous tone recognition system has been established by hidden Markov model. Experiment results show that correct recognition rates of system are 74.31% and 71.37% in the training set and testing set respectively. But for tone-three, the correct recognition rate is lower.For the characteristics of tone-three having low recognition rate, specific context of the tone-three syllable has been studied in the paper. The results show that the correct recognition rate is about 80% when tone-three syllable is in the end of sentence or word. Taking into account of the syllable's context and tone characteristic, tone recognition system based on context has been built by adding the sandhi rules into the system. Then the recognition rates of system have been improved respectively by 24.5% and 21.1% in training set and testing set, especially recognition rate of tone-three has been enhanced significantly. The results of the experiment indicate that the performance of tone recognition system based on context has been improved.In this paper, Chinese continuous digit string recognition system has been studied at last. It can be found that some number pairs are easy to misjudge,7 judged as 4 easily and 6 judged as 9 easily. For the tone having the role of meaning recognition, continuous digit string recognition system based on tone information has been built by adding tone recognition module into the system. Experimental results show that the correct recognition rates of improved system are 88.62% and 83.36% in the training set and test set respectively, and the false number pairs have got rectified significantly. The performance of system has been enhanced by adding tone information into continuous digit string recognition technology.