基于HMM-UBM和短语音的说话人身份确认

Research on the HMM-UBM and Short Text Based Speaker Verification

摘要: 提出了一种简单有效的与说话人有关的阈值设置方法，同时在确认系统中采用高斯混合模型（GMM）作为背景模型来模拟未知的冒认者的语音，在冒认文本多样化的情况下获得了比用隐马尔可夫模型（HMM）作背景模型更好的性能．本文在此基础上实现了一个基于Internet的与文本有关的远程语音身份认证系统，在实验测试和应用测试中均获得了满意的确认性能．

Abstract: In this paper,a simple and effective threshold-setting method for text-dependent speaker verification is proposed.The Gaussian mixture model(GMM)is also used as the background model to simulate the speech of unpredicted impostors and outperforms the conventional HMM(hidden Markov model)background model in the case of diversified pseudo text.Based on these two technologies,aremote text-dependent speaker verification system over Internet is proposed.This system is evaluated in both experimental and real condition and shows satisfied verification performance.