The Research And Implementation Of Continuous Speech Recognition System Based On Word Network Model

As computer science and information technology develop, speech recognition will become the main communication tool between human and machine. Through its development in about half a century, speech recognition technology has become more and more perfect and there are more and more products that are changing people's life.Speech recognition system has been widely used in various fields, such as voice dialing, voice commands, voice menu, voice navigation and so on. This paper research the key technology of continuous speech recognition. Design and develop a word network continuous speech recognition system-"MYASR". It is easy to build an word network continuous speech recognition application system on MYASR. As a experiment platform, all experiments in this paper are performed on MYASR.To improve the robustness of the voice activity detection (VAD), A GMM-based VAD approach has been proposed in this paper. With this method, two GMMs are constructed to model the noise and the speech respectively in the spectrum feature space, and signal frames are recognized by computing and comparing the outputs of both GMMs. Also, we have designed a method to self-adapt the GMM parameters according to various environmental conditions. Experimental results show that the proposed method generally performs better than traditional approaches such as the method based on short-term energy in noisy environments.Common speech recognition system systems are usually based on continuous density HMMs, which are typically implemented using Gaussian mixture distributions. Such statistical modeling systems tend to operate slower than real-time. To solve this problem, approximate likelihood evaluation methods are approached. Experiment showed that through approximate likelihood evaluation methods the decoding time decrease by more than 10%. But the traditional approximate methods produce likelihood deviation. By researching on various approximate methods, a new method TDGS based on DGS is approached. By comparing the result of phone recognition on Timit, TDGS showed great performance.