Automatic Transcription System for Lectures and Seminars

This technical report presents research conducted on developing an automatic transcription system for lectures and seminars as part of the FP6 Integrated Project Chil. The study utilizes widely available corpora to train acoustic and language models due to the limited data from the Chil project. The speech recognizer is based on technology from the Limsi Broadcast News Transcription system, employing various audio corpora for training. The report details the experimental setup, including the use of five seminars with diverse accents for testing and development. Results indicate a significant reduction in word error rates when appropriate acoustic and language models are applied. The report also discusses the implementation of a neural network language model, which further enhances transcription accuracy. Challenges in transcribing spontaneous and far-field speech are acknowledged, with ongoing efforts aimed at improving recognition performance. The findings underscore the complexities involved in automatic transcription tasks and the potential for future advancements.