Based On The Sound Of Vowels Chinese Voice Recognition System

Speech recognition is one of important technologies for speech communication between man and computers, which are in good prospects. Large-vocabulary Chinese speech recognition, as a significant branch in speech recognition is worthwhile to research because of its practical application.This paper studies some technologies, which mainly exploit INITIAL/FINAL units as recognition unit for Large-vocabulary recognition. A speech analysis system is also programmed. The tasks can be divided into three prime sections as followings:(1) A system of Isolated Mandarin Speech Recognition of a moderate range of word vocabulary(500 words)based on the unit of INITIAL/FINAL has been established. "INITIAL" means the initial consonant of the syllable, while FINAL means the vowel or diphthong part but including possible medial or nasal ending. The main tasks are: choose proper INITIAL and FINAL units; split syllable into INITIAL units and FINAL units; study and realized algorithm of the INITIAL recognition; study and realized algorithm of the Mandarin Speech Recognition based on the unit of INITIAL/FINAL, in which both INITIAL recognition and FINAL recognition use the LPC(Linear Prediction Coefficients) cepstrum as the speech feature and the HMM as speech model.(2) A system of INITIAL recognition is established. INITIAL recognition experiments are for the INITIAL part of all Chinese syllables with special person. Experiments are done to discuss the most difficult part of syllable recognition based on the unit of INITIAL/FINAL-- INITIAL recognition. The length of INITIAL unit is counted by using 3730 INITIAL units. According to the results of the statistics, an algorithm of pre-selection is presented, which can reduce up to half of the computation, and the correct recognition in INITIAL units can be in an accuracy of 85.07%.(3) A system aiming at speech recognition has been realized, in which the function contains: real-time operation, temporal features analysis and spectral features analysis of WAVE file, speech recognition based on the unit of INITIAL/FINAL.