| As the main research area of intelligent computer, Continuous speech recognitionis a technology to change the continuous speech signal into corresponding textsequence correctly and effectively, has been a concern of the national scientificcommunity. Research on the continuous phoneme recognition has a certain value inautomatic identification of the language or dialect, auto speaker recognition, languagelearning and other fields.In this paper, HTK (Hidden Markov model-based toolkit) based Uyghurcontinuous phoneme recognition baseline system is presented, conducted research andanalysis on HTK (based on the Hidden Markov Model Toolkit), on this basis, firstresearched the feature extraction method of its language-related aspects, acousticmodel and language model and other key technologies are addressed. According tothe characteristics of Uyghur language, designed the text corpus for languagemodeling and speech corpus construction, and recorded a large-scale speech data,determine the phoneme list, designed the context attribute, problem sets and decisiontree, and finished the necessary works to establish the database. Configured UyghurHMM topology for training the Uyghur monophone and triphone based Uyghuracoustic model. The different recognition rates of the monophone and triphone basedacoustic models under the different language models and Gaussian mixture are alsogiven in this paper. The statistics of the recognition rates of32Uyghur phonemes, thelist of the confused phonemes and their possible reasons are analyzed. Established anUyghur continuous phoneme recognition experimental platform, and laid thefoundation to further improve the recognition rate.This research not only can be used in the phoneme recognition, can also beapplied to other Uyghur speech processing, and provided the necessary preparatorywork and some reference for others research. In addition, the research work has ahigher reference value for the entire Altaic languages, such as Kazak and Kirgiz. |
PDF Full Text Request