Chinese Speech Pretreatment And Isolated Word Recognition Method

As a new cross subject, the speech recognition can be very precious for scientists to do research and make applications . The technology of speech recognition and speech synthesis has been landmarks of modern technology, and it is also an important field of modern computer technology. Although some breakthrough had been made and some products were proposed, most of the speech recognition systems are still kept in lab because of the limitation of accuracy and complexity.The isolated speech recognition system for small vocabulary is important and widely used. In the application of rehabilitation exercises for hearing impaired children, the features of the speech of the children, mostly isolated words, are analyzed and compared with those of standard speech. The trainer can correct speech of themselves by the results of the comparisons. In this, the extraction, analysis and recognitions of the features are language-independent. This paper is part of a vision-based computer-aid training system for Chinese hearing impaired children. The isolated speech recognition system was studied in detail in this paper.The pre-process of speech signal was first analyzed and an improvement was made to the endpoint detection method in which the speech was mapped to a voice matrix which was based on a model called mind acoustics. Then the sum of voice matrix that can be adapted to different background noise circumstance was got. Connecting with the characteristic of Chinese speech, speech endpoint can be detected well without changing the threshold.And then the isolated speech recognition system based on dynamic time warping (DTW) was studied. A real-time speech recognition system was presented based on a combination scores. In the process of recognition, two improved DTWs were attained by vector-interpolation. The two improved DTW recognizers were then combined with original DTW recognizers to recognize speech .Tests showed that the system can achieve higher recognition rate than DTW-based recognition system in low SNR environments.Isolated speech recognition system based on HMM was also studied in this paper and the problems of optimization calculation were analyzed intensively. The revaluation formula of parameters without underflow was given. The speaker-independent DHMM speech recognition system based on Corrective Training was studied and applied to speaker-independent CHMM speech recognition .The results of experiments showed that the recognition rate was improved.