Researching And Building Of The Mongolian Large Vocabulary Independent Continuous Speech Recognition System

To make computer hearing is the object of speech recognition, namely make the computer exactly recognize the content of speech under various conditions, and can implement the intention of people according to the speech. Speech recognition is a cross-subject relating to many aspects, which has close relationship with computer science, communications, linguistics, signal processing and artificial intelligence. Mongolian is an influential national language in the world, which has been widely used in the minority regions such as Inner Mongolia Autonomous Region. Under the background that English and Chinese speech recognition system have already gone out of laboratory and had been widely used, however there is not a system about Mongolian speech recognition, we choose the Mongolian speech recognition system as researching subject is significant to make for prosperity of minority's culture and social progress of minority district.Speech recognition system mainly includes isolated word recognition, continued word recognition and continuous speech recognition system, as well as small vocabulary, medium vocabulary and large vocabulary speaker dependent and speaker independent recognition system. The speaker dependent small and medium vocabulary recognition system is easily to build and can achieve relatively high recognition rate. To build a speaker independent large vocabulary speech recognition system is very difficult, and its recognition rate is relatively lower than speaker dependent small and medium vocabulary recognition system. But the system is universal and can adapt many instances, and it is the primary research direction in recent years.The main research approach in speech recognition is stochastic model; the excellent representative of the approach is speaker independent large vocabulary speech recognition system based on HMM (Hidden Markov Model). The approach uses tri-phone models as recognition sub-word unit in the acoustics and speech layer, and uses statistic language model in syntax layer which is based on bi-gram and tri-gram.The Mongolian large vocabulary speech recognition system uses HTK as training and recognition toolkit, uses tri-phone models as basic recognition sub-word units, and experiment on formula based language model as well as statistic language model.