Using textual information from LVCSR transcripts for phonetic-based spoken term detection

This paper presents a spoken term detection method, based on automatic speech recognition and phonetic representation. The proposed method combines textual search in word transcripts obtained with a large vocabulary continuous speech recognizer system and phonetic search in the phonetization of these transcripts, to accurately locate the occurrences of a list of keywords in a broadcast corpus. Textual information from the transcripts and an efficient rescoring scheme are used to improve the performance of the phonetic search. Our experiments show that the proposed method outperforms the baseline textual and phonetic searches by its ability to separate correct detections from false alarms.