基于GMM模型的声纹识别模式匹配研究

模式匹配是声纹识别的关键问题之一,为了提高识别正确率和识别效率,本文采用GMM模型建模,训练阶段利用EM算法求取参数集,并通过MAP准则实现模式识别。引入LBG算法求取起始参数值,并设计了基于3种方法的联合判决门限决策。实验结果表明 GMM模型利用平均值向量和协方差矩阵使它具有更好的模型能力,当高斯混合数为32时识别率达到最高,联合判决门限决策有效降低了误识率与虚警率,并提高了识别效率。

Pattern Matching of Voiceprint Recognition based on GMM

Pattern matching is one of the key problems of voiceprint recognition. In order to improve the accuracy and efficiency of recognition, this paper adopts GMM modeling, applies the EM algorithm to calculate parameter set during the training stage, and achieves pattern recognition via MAP criterion. LBG algorithm is introduced to calculate the initial parameter values, and a combined threshold decision is designed based on 3 methods. Experiment results show that GMM, with mean vector and covariance matrix, enjoys better modeling capability, and reaches the highest recognition rate when the mixed number is 32. The combined threshold decision effectively reduces the false acceptation rate and false alarm rate, and meanwhile, it improves the efficiency of recognition.