The Research Of Speaker Recognition Based On Time-frequency Feature

Speaker recognition is an authentication of biological to identify speaker according to the voice, widely used in speech retrieval, judicial identification, computer remote login, safety certificate, etc. Because of its wide application, the speaker recognition has become a hot spot of research in the field of biological anthentication technology today. The speaker recognition attention is paid to the speech signal in the speaker's personality factors, emphasized the difference between different people, which requires the speaker feature extraction to very precise. Although there have been various extraction method, there are some difficulties by duing to the environment and speak in one's own various factors. This paper focus on the speaker's time-frequency feature, and the main research works are as follow:1. This paper puts forward a time-frequency method based on Gabor transform and bilinear time-frequency distribution of combining. According to the short-time Fourier transform, Wigner-Ville distribution, Choi-Williams such as distribution of time-frequency analysis problems, from the Angle of image processing of Gabor transform and bilinear time-frequency distribution of time-frequency combination after image processing technology, can get good time-frequency gathered sex also reduced cross terms of time-frequency distribution. Simmulation results which are compared with Gabor transform, Wigner-Ville distribution and Choi-Williams distribution it can gain finer time-frequency structure.2. Through the front speech time-frequency structure, using Mel-Frequency Cepstrum Coefficients as the characteristic parameters for speaker recognition. First, respectively extractint the Mel frequency cepstrum parameters of primitive voice, various respectively on time-frequency analysis methods and the improved time-frequency structure, and then use support vector machine for speaker recognition.Simulation experiments show that the improved time-frequency structure is more useful to improve speaker recognition, and recognition rate increased 94.1%.3. Through the analysis about the human ear auditory system analysis, we have proposed using Gamma tone filter to speaker of processing methods. First,we studied pulse coupled neural network theory and two commonly used features, namely, time series and entropy sequence, and compare the two features for each speaker recognition.The simulation results show that the entropy sequence more can be a very good recognise different speaker. Then the speaker's voice signal through Gamma tone filter, and then put processed signals through spectrogram algorithm and get spectra.Finally,we can extract feature by using pulse coupled neural network, obtain each speaker of entropy sequence and compare with different speaker. Simulation results show that the Euclidean distance of the spectrogram's entropy sequences is smaller than not through the Gamma tone filter smaller.