Technology Of Tibetan Speech Recognition Based On Fast Walsh Transform

The research of Tibetan speech recognition is still in the initial stage due to the various reasons. Because there is a large population use Tibetan and development of the technology can promote academic exchange and connection between Tibet and outside world what makes great sense in promoting national unity and stability. So it plays extremely important role in wide range applications of speech identification systems.Identification speed is one of the most prime targets for the recognition of isolated Tibetan speech. But it will be unable to meet the real-time requirement when the vocabulary increases greatly. The fast Walsh transform was applied to extract the feature parameters instead of MFCC to solve this problem. It shortens the duration of parameters calculation and the systems improved obviously.For the recognition of continuous Tibetan speech, the precision of dividing the continuous speech into units determines the recognition effect. MFCC_FWT and screened twice based on Wavelet Transform segmentation algorithm were applied to process the continuous speech of Tibetan. So the continuous speech was divided into units and it could be identified.The main work and contributions are as follows:1. Analysis the Tibetan pronunciation features and sentences'syntactic features firstly. Introduce the basic principles of Tibetan speech recognition system. In-depth presented the principles of pre-processing and endpoint detection technology.2. Introduced the algorithm of MFCC. Introduce the fast Walsh transform into it to improve the computational efficiency. After the improvement, the computed speed increased greatly. And it can also ensure the effectiveness of characteristic parameters.3. Present the both DTW and HMM recognition algorithms and applied them into the isolated word Tibetan speech recognition system of medium vocabulary. DTW algorithm is rather simple and effective for isolated word recognition of a particular person. HMM algorithm has strong modeling capability, so it can conveniently represent any voice-based element. HMM is suitable for both isolated word and continuous speech recognition systems.4. MFCC FWT and screened twice based on Wavelet Transform were applied to the segmentation algorithm of continuous speech in Tibetan for the first time. Then recognize the continuous speech after it was divided into the isolated units. It greatly simplifies the system's realization difficulities for continuous Tibetan speech.