The Cambridge University spoken document retrieval system

TLDR

The paper presents a spoken document retrieval system and evaluates its performance on approximately 50 hours of automatically transcribed broadcast news. The system combines an HTK-based speech recognizer with a City University–derived retrieval engine, and is tested on the 1998 TREC‑7 task and on varied transcription error rates, with new error metrics defined. Results demonstrate that higher transcription accuracy markedly improves retrieval performance, and the newly defined metrics more accurately reflect the impact of transcription errors.

Abstract

This paper describes the spoken document retrieval system that we have been developing and assesses its performance using automatic transcriptions of about 50 hours of broadcast news data. The recognition engine is based on the HTK broadcast news transcription system and the retrieval engine is based on the techniques developed at City University. The retrieval performance over a wide range of speech transcription error rates is presented and a number of recognition error metrics that more accurately reflect the impact of transcription errors on retrieval accuracy are defined and computed. The results demonstrate the importance of high accuracy automatic transcription. The final system is currently being evaluated on the 1998 TREC-7 spoken document retrieval task.