A novel approach for spoken term detection in Vietnamese

In this paper, we describe our results on spoken term detection (STD) for Vietnamese speech. A method for constructing the indexing module and the searching module for Vietnamese speech is presented. A new method for identifying detection threshold is also described. The experiments were carried out with different thresholds: normalized acoustic probability and posterior probability. For our results, optimal Actual Term-Weighted Value (ATWV) is 0.5982 (the probability of false alarm is 0.48% and the probability of miss detection is 20.85%), corresponding with the test using normalized posterior probability as detection threshold.