节点文献
跨信道说话人识别语音库的设计与研究
The Design and Research of Cross-channel Speaker Recognition Database
【Author】 Lin Li;Lihong Wan;Ling Huang;Qingyang Hong;School of Information Science and Technology, Xiamen University;
【机构】 厦门大学信息科学与技术学院;
【摘要】 建立一个适合于特定说话人识别系统的汉语语音数据库,对推动说话人识别技术的研究和应用具有重要意义。本文针对说话人识别中信道差异问题,设计并建立了跨信道说话人识别语音库XMU-Multi Channel。本文首先分析比较了国内外现有的说话人识别语音库,设计了跨信道说话人识别语音库的语料内容、采集系统、录音方案以及标注和存储方式,进一步基于这个数据库实现了PLDA参数更新的跨域迁移策略,以解决小样本信道PLDA建模的困难,并在语音样本有限的前提下提高了说话人识别系统的识别性能。
【Abstract】 The establishment of a Chinese speech database has a great significance in promoting the research and application of speaker recognition technology. This paper designs a cross-channel speech database XMU-Multi Channel for the research of channel mismatch in speaker recognition. After comparison of existing domestic and international speaker recognition speech database, this paper has a discussion on recording program, acquisition system, recording content, labeling and storage. Based on our designed XMU-Multi Channel database, a cross-domain transfer strategy experiment of PLDA was implemented. As a result, the difficulties for small samples to train the PLDA model could be resolved.
【Key words】 speech database; speaker recognition; cross-channel; cross-domain transfer;
- 【会议录名称】 第十三届全国人机语音通讯学术会议(NCMMSC2015)论文集
- 【会议名称】第十三届全国人机语音通讯学术会议(NCMMSC2015)
- 【会议时间】2015-10-25
- 【会议地点】中国天津
- 【分类号】TN912.34
- 【主办单位】中国中文信息学会语音信息专业委员会