Research Of Speech Recognition Based On Telephone Channel

Speech recognition has been widely adopted in many aspects of communications, not only along with the growth of the ambulant requirement for people to obtain information but also the widespread of telephone networks. So the application of speech recognition in the telecommunication field has become one of the most important trends. However, how to improve the performance of speech recognition system under telephone environment has been a great challenge. This paper studies from feature space to model space, in order to improve the performance of telephone speech recognition system.Firstly, the paper expatiates on the nonlinear and chaotic characters of the speech signal and explains the nonlinear academic foundation of speech processing techniques. Then a basic method to extract the nonlinear feature for speech signal is illuminate, which is named reconstructed phase space In order to create phase space, the approach to determine time lag and embedding dimension is studied. A new method called combine average displacement is presented for improving the average displacement. Then the combination of both RPS and MFCC feature sets into one joint feature vector is examined. Then a new feature RPSMECC is presented. At last, based on detailed comprehension in the influence of telephone channel and the time-based character of RPS feature, the paper proposes a new method combined the neural networks adaptive noise cancellation with RASTA.Later, the conception of hidden markov model to speech recognition is studied deeply as well as universal background model. Then a new model incorporated hidden markov model with universal background model is proposed.Lastly, a telephone speech recognition system is presented using Matlabprogramming language. Then the proposed feature parameters, model and training algorithm are experimented by the system. The results indicate that the recognition system has achieved a higher accuracy.