In this paper, the general computation methods of confidence measure (CM) and its applications are discussed under different grammar constraints. Here confidence measure is divided into two parts: acoustic CM and linguistic CM. By combining these two parts, a unified computation algorithm of confidence measure is proposed, where the detailed algorithms and their applications are explored under different grammar constraints respectively. The contribution of the thesis is shown as follows: (1) Firstly, confidence measure is divided into two parts, whose general computa- tional methods are given, which makes it possible to cast different grammar- constraint condition into one unified framework. (2) Secondly, taken CTS keyword spotting system as an example, the confidence measure algorithm under weak-sense grammar constraint is studied, and MCE optimized acoustic confidence measure and context enhanced verification method are introduced, which makes use of discriminative training and local linguistic information to get better performance. In CTS keyword spotting system, EER drops by 13.8% relatively. (3) Thirdly, taken LVCSR system for instance, statistical grammar-based confidence measure is explored, and word graph posterior probability based CM can be regarded as linguistic confidence measure, which indicates that it is similar to online garbage model based CM. On 2004 “863” national LVCSR evaluation set, the EER of confidence measure is 22.7%. (4) Finally, in key slot spotting system, confidence measure under strong-sense grammar constraint is explored, which shows how dynamic extension of slot grammar will influence the compression of search space and suppression of search error, and CM based pruning method can reduce the search error as well. By these methods, the slot accuracy of the system increase from 47.1% to 65.2%.
修改评论