The Research Of Speaker Recognition Algorithms Based On MFC And Vector Quantization

Speaker recognition as one of the biometrics techniques is to recognize speaker’sidentity automatically from its voice waveforms which contains differences ofphysiological and behavioral characteristics specific to each individual. Speakerrecognition has caught many people’s attention for its wide application prospect. Thisthesis focuses on the research of text-independent speaker recognition technologybased on MEL frequency cepstral coefficients (MFCC) and vector quantization (VQ).Paper combines VQ with adaptive ant colony algorithm and uses improved MFCC asits parameters.First, we improve the window function by adding the new Lanczos windowfunction in MFCC extraction preprocessing instead of the Hamming window. With thewidth of lobe is almost the same, we adopt a function that more sidelobe suppression.We also introduce Bark wavelet transform (BWT) for more suitable to human ear’sauditory system, the base function of BWT obeys the time-frequency optimaluncertainty but the scale function varies according to the critical band.It can improvethe speech’s robustness in noisy environment in MFCC feature extraction process.Then, improve the pattern method of speaker recognition base on vectorquantization.To make up limitation of selection of initial codebook and avoid thepossibility that classical LBG algorithm easily obtain local optimization, we adopt thethe method that combine of VQ and ant colony algorithm (AC-LBG) with distributedparallel mechanism, we mix them and alternate them.This can not only improve theglobal search ability, but also improve convergence speed with LBG algorithm. Inorder to prevent the ant colony algorithm falling into local optimal solution, we adopta combination of the certainty selection strategy and random selection strategy in thesearch process, and the strategy can adjust transition probability dynamically, this canprevent the obtained result from local optimal solution in a certain extent.Finally, we present a weighted Euclid’s distance distortion measure based on thestandard deviation in the matching decision.From the experiments we can know that the recognition performance whichadopts the improved MFCC and AC-LBG is better than the traditional algorithm, andthe longer training speech length is, the better recognition performance is. So we must adjust the number of codeword in practice.