Skip to Main Content
This paper presents an improvement for confidence measure estimation as posterior probabilities on lattices in speech recognition. An observation is presented that nontarget regions, i.e. non-speech part of a spoken utterance, of different lengths may lead to different levels of over optimistic confidence measures. This may be problematic in obtaining a consistent rejection performance at the same confidence threshold. Therefore an improvement is proposed to exclude the non-target region from the confidence measure (CM) evaluation procedure. Results of a set of experiments demonstrate effectiveness of the proposed procedure and achieve and finally a uniform rejection performance can be achieved at fixed thresholds.