A Combination of Maximum Likelihood Bayesian Framework and Discriminative Linear Transforms for Speaker Adaptation

摘要：

Linear transforms are one of the most commonly used methods to speaker adaptation. In this paper, we present a combinational method of Bayesian framework and maximum likelihood linear regression as well as discriminative method for speaker adaptation. Furthermore significant gains can be obtained using discriminative training for acoustic models. Experiments on supervised adaptation on Persian data show that the combinational method outperforms both Maximum likelihood linear regression and Bayesian framework. Also the proposed method with discriminative adaptation outperforms previously proposed methods for transform estimation and discriminative training outperforms ML training.

关键词： Speech recognitlon speaker adaptation maximum-a-posterior adaptation maximum likelihood linear regression adaptation discriminative linear transforms

作者: Shadi Pirhosseinloo Shahram Javadi

作者单位: Scientific Association of Electrical & Electronic Engineering Islamic Azad University Central Tehran Electrical Engineering Department Islamic Azad University Central Tehran Branch TEHRAN, IRAN

会议类型: 国际会议

会议名称: 2011 3rd International Conference on Computer and Automation Engineering(ICCAE 2011)(2011年第三届IEEE计算机与自动化工程国际会议)

会议地点: 重庆

会议语种:英文

页码: 107-110

在线出版日期: 2011-01-21（万方平台首次上网日期，不代表论文的发表时间）