Introduction of false detection control parameters in spoken term detection

This paper describes spoken term detection (STD) with false detection control. Our STD method uses phoneme transition network (PTN) derived by multiple automatic speech recognizers (ASRs) as an index. An PTN is almost the same to a sub-word based confusion network (CN), which is derived from an output of an ASR. The PTN-based index we proposed is made of the outputs of multiple ASRs, which is known to be robust to certain recognition errors and the out-of-vocabulary problem. Our PTN was very effective at detecting query terms. However, the PTN generates a lot of false detections especially for short query terms. Therefore, we applied two false detection control parameters to the Dynamic Time Warping-based term detection engine. In addition, we changed the search parameters depending on length of a query term. Finally, the STD performance was better (0.785 of F-measure) than without any parameters (0.717).