Phonetic Spoken Term Detection in Large Audio Archive Using the WFST Framework

The paper presents a technique for phonetic spoken term detection in large audio archive. It is designed within the framework of weighted finite-state transducers and utilizes the rather recently developed notion of factor automata, which we have enhanced with a score normalization and a technique for systematic query expansion which allows for phone deletions and substitutions and consequently compensates for frequent pronunciation imperfections and systematic phoneme interchanges occurring during the ASR decoding process. The experiments presented in the paper show that the new WFST-based method outperforms the baseline system both in terms of search performance and speed. Finally, the paper discusses the issues of the proposed techniques that need to be addressed before the application in real-life tasks.