Combining State-Level Spotting and Posterior-Based Acoustic Match for Improved Query-by-Example Spoken Term Detection

In spoken term detection (STD) systems, automatic speech recognition (ASR) frontend is often employed for its reasonable accuracy and efficiency. However, out-of-vocabulary (OOV) problem at ASR stage has a great impact on the STD performance for spoken query. In this paper, we propose combining feature-based acoustic match which is often employed in the STD systems for low resource languages, along with the other ASR-derived features. First, automatic transcripts for spoken document and spoken query are decomposed into corresponding acoustic model state sequences and used for spotting plausible speech segments. Second, DTW-based acoustic match between the query and candidate segment is performed using the posterior features derived from a monophone-state DNN. Finally, an integrated score is obtained by a logistic regression model, which is trained with a large spoken document and automatically generated spoken queries as development data. The experimental results on NTCIR-12 SpokenQuery&Doc-2 task showed that the proposed method significantly outperforms the baseline systems which use the subword-level or state-level spotting alone. Also, our universal scoring model trained with a separate set of development data could achieve the best STD performance, and showed the effectiveness of additional ASR-derived features regarding the confidence measure and query length.