Pruning Strategies for Partial Search in Spoken Term Detection
Van Tung Pham, Haihua Xu, Xiong Xiao, Nancy F. Chen, Eng Siong Chng · 2017
In this paper, we propose a partial search approach for subword-based keyword search (KWS) systems. The proposed approach addresses the problem of high miss rate in the conventional full sequence matching approach by retaining detections that only contain some partial sequences. To control the increase of false alarm (FA), we propose two pruning methods, global threshold pruning (GTP) and keyword-specific pruning (KSP). The former uses the same threshold for all keywords; while the latter gradually adjusts a specific threshold for each keyword. Experimental results on the NIST OpenKWS15 and OpenKWS16 Evaluations show that partial search using KSP is effective. For out-of-vocabulary keywords, our proposed approach reduces the miss rate up to 21.0% absolute and improves the Actual Term Weighted Value (ATWV) up to 19.2% absolute over the full search baseline