Improved spoken term detection by template-based confidence measure

The confidence measure (CM) in spoken term detection is addressed in this paper. We present a template-based approach in CM procedure. The template speeches are first selected from the development corpus. Then the Dynamic Time Warping (DTW) algorithm is adopted to calculate the distances between the candidate and template speeches. These DTW distances are transformed to the confidence score through some algebra processing. Furthermore, the DTW scores and posterior probability from the speech recognition lattice can be combined to improve the term detection. Experimental results show that the proposed approach offers significant performance improvement. A 22.3% relative increase in Average Term Weighted Value (ATWV) can be obtained in the evaluation set.