Design and Implementation of a Real-Time Speaker Identification System with Improved GMM

The text-independent real-time speaker identification system is presented. It is based on Gaussian Mixture Model and MFCC (Mel frequency cepstral coefficients) method to extract the character of speech signal. The traditional method of GMM parameters initialization includes random method and k-means clustering are lack of clustering accuracy. A new approach which combines division and k-means clustering is presented and applied to the system. The system is realized under windows platform with good face. It includes voice collection and storage, speech pre-processing, MFCC extraction, GMM training and storage, speaker identification and so on. The experiment shows that the improved method as compared with the traditional method, the system average recognition rate has an increase of 18.34% and 7.98%. The system can achieve the error rate with 6.7% under the provided experimental condition.