关键词: 高斯混合模型, 语音活动检测, 识别错误率

Abstract: In order to investigate the function of Gaussian Mixture Model（GMM） in speaker recognition，a GMM based speaker recognition system is designed.The system consists of four modules that are audio signal pre-processing，speech activity detection，speaker modeling as well as audio signal recognition.The first three modules constitute the model training segment of the system and the last module constitutes the speech recognition segment of the system.A speech activity detector which is built by GMM in the second module is the innovation of the research.Some tunable parameters and recognition error rate of the system are tested using audio-visual meetings in the Augmented Multi-party Interaction（AMI） corpus.Simulations show that with the help of the speech activity detector and several filter algorithms，recognition accuracy rate of the system for audio signal with overlap speech can reach 83.02%.

Key words: Gaussian Mixture Model（GMM）, speech activity detection, recognition error rate