Research Of Recognition And Transformation Of Speech Signal

There is long time of speech signal processing and it presents the extensive practical field. In the thesis the author presents some research on speaker recognition, mixed speech signal separation and speech transformation using neural network..There are six chapters in the thesis. The first chapter is the background of the research. In chapter 2, the content of speech signal analysis is introduced. Chapter 3 describes the content of speaker recognition and the abstraction of the characteristic parameters. Chapter 4 describes the Hidden Markov Model completing the recognition of speaker. Chapter 5 introduces the theory and method of the separation of the linear mixing speech signal based on blind signal separation. Chapter 6 introduces method to speech transformation using Neural Network and some experiment results. At last, we draw some conclusions.Through the theory analysis and experiment simulating, we get some acquaintance and conclusions as follows:(1) The abstraction of the characteristic parameters is the important part in speaker recognition. It influences the effect of recognition. In the thesis, we select the Mel-Frequency Cepstrum Coefficients based on analyzing a lot of parameters of speech signal. Mel cepstrum is of better recognition and anti-noise capability.(2 ) Dynamic Time Warping, Vector Quantization, Hidden Markov Model and Artificial Neural Network can be used in speaker recognition. HMM is the best model of speaker recognition and it is applied in the thesis, which is the model from left to right, and recognizes the two speakers.(3) The speech signals are mixed signals in life. It is useful to recognize speaker using blind separation technology, which can be used to separate the mixed signals. We separate the linear mixed signals using the fourth order cumulants and Independent Component Analysis technology. The source signals, which are separated, are applied to speaker recognition and they can be well recognized.(4) In the thesis, the transformation of speech signals of different speaker is completed by BP Neural Network. The transformation of single word is completed.