Classification of speech signals is an important base of the speech recognition,speaker recognition and speech synthesis, for which the chief issue is to select the propercharacteristics of signals.Firstly, we reviewed the research of time-frequency analysis and speech signalprocessing. Then the dissertation studies the principles, applications and characteristic ofthe speech characteristic representation technologies, such as Short Time FourierTransform(STFT), Liner Prediction Coding(LPC), Gabor Transform and WaveletTransform, and arithmetic model, such as Minimum Distance Classifier, VectorQuantization(VQ), Gaussian Mixture Model(GMM) and Hidden Markov Model(HMM). Since STFT and Wavelet Transform can't represent the fine structure ofspeech signal, the thesis uses quadratic time-frequency distribution as the feature of thesignals. To obtain the better performance of classification, the selected kernels areoptimized by Nelder-Mead arithmetic. Finally, the experiments of speech recognitionand speaker recognition shows that we can obtain the perfect performance ofclassification when quadratic time-frequency distributions with optimized kernel areused as characteristics. |