Speech Recognition Of Connected Mandarin Digits

Continuous digit speech is one of the greatest challenges in Mandarin speech recognition. Reviewing the state and the art of Mandarin digit speech recognition, the paper analyses the high confusion difficulties in Mandarin digit speech recognition, the methods of Continuous Digit Recognition are studied. Digit speech recognition system, which uses MFCC as feature parameters and uses weighted method to analyze speech feature, considering different contributions of features. With a speech simulating system, performances of these two characteristics are compared, and the noise robust property of MFCC is investigated. The problems and challenges for improving robust Mandarin digit speech recognition at the acoustic processing stage was also discussed. This paper presents our work on endpoint detection based on mel–scale features and phoneme segmentation. The conventional speech detection methods based on some simple features such as energy cannot work as good as mel–scale features. The experiments show that the high accurate detection can be obtained. The experiments show that the method can better reflect the differences of vowel and consonant. The system effectively eliminates the deviation caused by background noise. In this paper, a combined model for speech recognition is proposed, based on VQ and traditional HMM. The efficient algorithm of train and recognition is proposed. in addition, easy implementation and low cost are preserved. The resulting performance of the connected mandarin digit recognizer show: ①the recognition speed is fast, but the requirement of hardware is low. The implement of recognition process can be done in PC; ②to isolated Mandarin digit pronounce , the correct rate is high; ③in the condition of continuous Mandarin digit speech, the correct rate does not lapse obviously.