Score distribution based term specific thresholding for spoken term detection

Abstract

The spoken term detection (STD) task aims to return relevant segments from a spoken archive that contain the query terms. This paper focuses on the decision stage of an STD system. We propose a term specific thresholding (TST) method that uses per query posterior score distributions. The STD system described in this paper indexes word-level lattices produced by an LVCSR system using Weighted Finite State Transducers (WFSTs). The target application is a sign dictionary where precision is more important than recall. Experiments compare the performance of different thresholding techniques. The proposed approach increases the maximum precision attainable by the system.