Abstract

In this paper we describe the IBM Audio-Indexing System which is a combination of a large vocabulary speech recognizer and a text-based information retrieval system. Our speech recognizer was used to produce the baseline transcripts for the NIST SDR97 evaluation. We report the performance of the system on the SDR-97 "known item retrieval" task and on a more pertinent TREC-style Audio-Indexing task. 1. Introduction The goal of an audio-indexing system is to provide the capability of searching and browsing through audio content. The system is formed by integrating information retrieval methods with large vocabulary continuous speech recognition. A large vocabulary continuous speech recognition system is used to to produce time aligned transcripts of the speech. Information retrieval techniques are then employed on these recognized transcripts to identify locations in the text that are relevant to the search request. These locations with time alignments then specify regions of the speech...

Citations