基于因子分析的会话语音说话人识别研究

项目名称： 基于因子分析的会话语音说话人识别研究

项目编号： No.11504406

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 数理科学和化学

项目作者： 杨琳

作者单位： 中国科学院声学研究所

项目金额： 24万元

中文摘要： 无论在国家安全还是民用领域，实际应用中获取的语音数据大多是包含多人会话的录音，如电话对话、会议录音、网络聊天等，针对这种实际语音进行说话人身份确认或追踪的问题尤为重要。对这类问题的解决包括说话人分段聚类和说话人识别两个方面，本项目重点研究基于因子分析的说话人建模技术在说话人分段聚类和说话人识别中的应用，通过对少量因子的估计建立更准确的模型；基于说话人因子研究未知类别数目情况下的说话人聚类算法；通过研究对说话人因子的补偿和打分技术，解决聚类后短时语音说话人识别问题和对聚类结果的容错性。在此研究基础上，构建基于因子分析的会话语音说话人识别系统，推动说话人识别技术在实际应用中不断完善和发展。本研究的相关建模方法和聚类方法对图像分析、语音识别的其他领域也有借鉴意义。

中文关键词： 说话人识别；说话人聚类；因子分析；分段聚类；说话人日志

英文摘要： Whatever in the field of national security or for civil use, the recorded speech are mainly saved as the conversational format, for example television and meeting recordings, and the voice chats in internet. It is important to study on speaker identification and diarization for the conversational speech. This problem can be solved by speaker segmentation and speaker clustering. To address these issues, this project focus on the research of speaker modeling based on factor analysis, which estimates a small quantity of parameters. Also based on the speaker factor, we make a study of the speaker clustering with automatically estimating the number of clusters. Moreover in this work we try to solve the short-time speaker recognition and enhance the fault-tolerant ability by factor compensation and scoring. Based on the research achievement we are devoted to establish the framework of speaker recognition for conversational speech, in order to motivate the practical application of speaker recognition. This modeling method can be further extended to other areas such as speech recognition and image analysis.

英文关键词： speaker recognition;speaker cluster;factor analysis;segment and cluster;speaker diarization