基于快速DGMM的随机提示文本的话者确认系统

THE SYSTEM FOR TEXT RANDOM PROMPTED SPEAKER VERIFICATION BASED ON FAST DYNAMIC GAUSSIAN MIXTURE MODEL

研究了随机提示文本的话者确认技术中的几个关键技术，包括确认算法的训练和识别速度、话者确认文本和说话方式的选择、测试文本长度的选择、阈值的设定及话者语音的中长期变异的自适应算法．为提高训练和识别速度，该文提出了快速动态高斯混合话者模型，讨论了音素对话者确认系统的影响及测试文本长度对话者确认系统性能的影响．提出话者语音特性的中长期变异性的自适应增量学习的方法．同时文中详细地分析了一次和多次测试时话者的弃真率和取假率的关系，提出一种确定阈值的新方法．话者确认实验表明，快速的动态高斯模型的训练速度和识别速度比隐马尔柯夫模型快很多，并且两者的识别率相当，该文提出的阈值设定方法及话者语音特性的中长期变异性的自适应学习的方法十分有效．

This paper describes several key techniques for text random prompted speaker verification problem, including the training and test speed of the recognition algorithm, the selection of speaker text and speaking styles, the length of test text,the threshold setting and the adaptive algorithm for long term speaker characteristics variations. To increase the training and test speed, this paper proposes the fast dynamic gaussian mixture model(FDGMM) for speaker verification, discusses the influence of different phonemes, speaking styles and the length of test text on the speaker verification, introduces an adaptive algorithm for long term speaker characteristics variations, analyzes the relationships between false accept rate and false reject rate of one try and multiple tries and proposed an approach for threshold setting. The experimental results have shown that the speed of FDGMM is much faster than that of HMM and the recognition rates of the two methods are almost the same, the algorithms for threshold setting and the adaptive algorithms for long term speaker characteristics variations are very effective.