Large Vocabulary Continuous Speech Recognition Research Based On HTK

Exchanging information with the outside world is one of the fastest, most effective and widely used way of exchange. Speech recognition technology through a number of disciplines, and it will be involved in many areas. For example, statistics, physiology, acoustics, information theory and computer science, digital signal processing technology, applied psychology and pattern recognition theory, etc.. In recent years, the research on continuous speech recognition, a large vocabulary,which is a difficult task. Although large vocabulary continuous speech recognition system has made a series of achievements, but with its wide application, the system also showed its shortcomings. Especially in the large-scale application of the same pronunciation of different word recognition on the recognition accuracy of the defects and the noise environment can not be very good recognition of the difficulties. To this end, the deeper research of the system is of great significance and value.In this paper, we study the large vocabulary continuous speech recognition, and research is proposed aiming at the noise environment based on Mel frequency cepstral coefficient similarity of endpoint detection methods. Contrast triphone and tone modeling methods are excellent in the process of implementation of, and based on the similarity of MFCC speech endpoint detection technology further to recognize speech. The model construction by way of contrast respectively by single phonemes and phoneme modeling, the triphone models in establishing a full consideration of the issues related to the context, so this modeling recognition rate is better than single phoneme model. Finally, when adding different noises, the validity of the method is studied by studying the similarity of Mel’s MFCC.Finally, six groups of contrast experiments were done, and the recognition rates of 10dB, 0dB and-10dB were obtained in the traditional endpoint detection. The recognition rates were 43.04%,12.62%, and 6.07%, respectively. In the use of the Mel MFCC coefficient of the similarity of the endpoint detection are also added 10dB, 0dB and-10dB noise, respectively, the recognition rate of the sentence:45.06%,41.19% and 29.23%. With the decrease of the ratio of signal to noise, the correlation coefficient of MFCC is also slow down, but it can still get a better detection result.