Research On Voiceprint Recognition And Algorithm Of Patten Match

Voiceprint Recognition (VR) as one of the biometric identification research aims to identify living persons from their voice. It is useful in person authentication, forensics and speaker tracking, etc. Comparing to other biometric identification methods like fingerprint or face recognition, voiceprint recognition doesn't require expensive specialized equipments and are effective especially for remote identity verification.After studying the voiceprint recognition techniques already exist, this thesis constructs a recognition system based on Mel-scale Frequency Cepstral Coefficients (MFCC)-speech signal feature extraction method and Gaussian Mixture Model (GMM)~speaker mathematical model which both are popular and effective for voiceprint recognition. For speaker identification, Expectation Maximization Algorithm (EM) is adopted to train speaker dependent model, and afterwards recognize speaker according to Maximum a Posteriori Criterion (MAP). For speaker verification, in order to reduce Inner Speaker Variation and maintain Outer Speaker Variation, a speaker independent model is trained first, and then the model is adapted to speaker dependent model by using Bayesian Speaker Adaptation. In the step of pattern match, Log-Likelihood Ratio Detector is created by integrating speaker dependent model and Universal Background Model (UBM) trained in advance. Using this detector the score of speaker verification is computed.The most attention of the paper focuses on pattern matching scheme .Improvement of pattern recognition algorithms is presented especially in the field of speaker model training. A robust cluster algorithm called Fuzzy C-means Cluster Algorithm (FCM) is introduced to improve EM training process. At the end, experimental results show that the system constructed work well with high recognition rate-Speaker verification system can reach 4.7% Equal Error Rate (EER); FCM is better than other hard cluster algorithms in finding the initial points of the GMM and recognition rate improves 3.5% on average as a result.