Strategies for Model Training and Adaptation Based on Data Dependency Control
(0) In speech recognition systems, decoding or data evaluation processes are sometimes performed as a part of model estimation process. For example, when unsupervised adaptation is performed, parameters are adapted using decoding hypotheses generated by an initial model. In these model estimation frameworks, how to design dependency structures between data and models is an important issue, since it significantly affects recognition performance. This paper overviews model estimation methods that have been proposed from this viewpoint. These include efficient cross-validation (CV) based parameter tuning and structure optimization, domain adaptation based on selecting training subset, and iterative parameter estimation techniques that integrate CV into the process.
Takahiro Shinozaki Sadaoki Furui
Chiba University, Chiba, Japan Tokyo Institute of Technology, Tokyo, Japan
国际会议
2011亚太信号与信息处理协会年度峰会(APSIPAASC 2011)
西安
英文
1-6
2011-10-18（万方平台首次上网日期，不代表论文的发表时间）