Recovery from false rejection using statistical partial pattern trees for sentence verification

Abstract

In conversational speech recognition, recognizers are generally equipped with a keyword spotting capability to accommodate a variety of speaking styles. In addition, language model incorporation generally improves the recognition performance. In conversational speech keyword spotting, there are two types of errors, false alarm and false rejection. These two types of errors are not modeled in language models and therefore offset the contribution of the language models. This paper describes a partial pattern tree (PPT) to model the partial grammatical rules of sentences resulting from recognition errors and ungrammatical sentences. Using the PPT and a proposed sentence-scoring algorithm, the false rejection errors can be recovered first. A sentence verification approach is then employed to re-rank and verify the recovered sentence hypotheses to give the results. A PPT merging algorithm is also proposed to reduce the number of partial patterns with similar syntactic structure and thus reduce the PPT tree size. An automatic call manager and an airline query system are implemented to assess the performance. The keyword error rates for these two systems using the proposed approach achieved 10.40% and 14.67%, respectively. The proposed method was compared with conventional approaches to show its superior performance.