Accurate speaker recognition based on adaptive Gaussian mixture model

An adaptive Gaussian mixture model (AGMM) with variable component numbers is proposed for accurate speaker recognition. According to the cluster property of speaker's acoustic feature distribution, an absorb-merge-split mechanism is utilized to adjust the Gaussian component numbers during the model training to overcome over-fitting and under-fitting in traditional models. Therefore every speaker's AGMM models have different distribution component number. The experiment result shows the recognition accuracy of the proposed AGMM is greatly improved compared with the traditional Gaussian Mixture Model (GMM). The error rates of recognition with MFCC and Bilinear frequency cepstrum coefficients (BFCC) decline by 41.41% and 22.21% respectively in relative.