Design And Implementation Of Phoneme Recognition

Phoneme recognition is a hot spot of research in artificial intelligence nowadays.Traditional phoneme recognizer is modeled by HMM-GMM model.And the phoneme recognizer of the CMU-SPHINX system proposed in this paper is based on HMM-GMM model.With the continuous reform and progress of technology in recent years,the Neural Network technology has made rapid development in the field of artificial intelligence.In view of the more expressive ability of characteristics in the neural networks and the great success which it has achieved,this paper propose another phoneme recognizer based on ANN,which is called LC-RC phoneme recognizer.Finally,we have made necessary improvements on the two base systems and the performance of the system has been improved.The main work of this paper is described as follows:(1)Implement CMU-SPHINX and LC-RC base phoneme recognizers relying on open source software,and complete the necessary analysis and test of the two base phoneme recognizers.(2)Implement the training and evaluation system of Chinese and English language model based on CMU-SPHINX system.And obtain the language model of the Lattice format in the use of HTK toolkit.(3)Study the phonetic composition of Chinese phonology.According to the characteristics of Chinese,reasonable reduce the number of Chinese phonemes from 125 to 38,which can greatly reduce the model complexity and has made some performances improvements.(4)Use pitch features to calibrate phoneme boundaries,providing support for LC-RC system data preparation.(5)For Chinese and English phonetic data,tracking changes in the loss function during neural network training of LC-RC system,and make the necessary optimization of learning rate and iteration situation according to the change of loss function,and has achieved some improvements.(6)On the basis of the base system,Combine the prosodic features of speech,which can improve the recognition accuracy of the two base system.