Research On The Acoustic Models And Implemention On The Keyword Recognition System

Recently, speech recognition techniques have been developing at a high speed and will be implemented in more and more fields in future. To let the machine understand what's the human being talking about, even to communicate with people is the biggest dream of us for a long time. Now most of the research in speech recognition is focused on the Large Vocabulary Continuous Speech Recognition System and achieves much improvement, but for the mandarin recognition, there are still lots of works to do with the accuracy and robust in the recognition system. It's still a tough task.This paper introduced the history and actuality of the speech recognition technology, and the basic theories of the speech recognition, including speech signal processing, acoustics feature extracting, HMM and so on. Further more, three main problems and main arithmetic are explained.Then the training corpus and environments of the mandarin digit acoustic models are displayed, also we discussed the characteristics of the mandarin digits and listed the training measures. After analyzed the difficulties and the defects of the models, we improved the models both from the labels side and the models side. We also got a conclusion and did the analysis from the experiments data.After that, the mandarin acoustic models are discussed. Then analyzed the Phone, the Syllable and the Initial/Final, and a keywords recognition system was imported. After analyzed the differences between the two different structures of the keywords recognition systems, the anti-models was trained and implemented in the keywords recognition system. The results of the anti-models were compared to the testing results of the competition set words. A conclusion about that was made in the end.