Transcribing Broadcast News: The LIMSI Nov96

In this paper we report on the LIMSI Nov96 Hub4 system for transcription of broadcast news shows. We describe the development work in moving from laboratory read speech data to realworld speech data in order to build a system for the ARPA Nov96 evaluation. Two main problems were addressed to deal with the continuous flow of inhomogenous data. These concern the varied acoustic nature of the signal (signal quality, environmental and transmission noise, music) and different linguistic styles (prepared and spontaneous speech on a wide range of topics, spoken by a large variety of speakers). The speech recognizer makes use of continuous density HMMs with Gaussian mixture for acoustic modeling and n-gram statistics estimated on large text corpora. The base acoustic models were trained on the WSJ0/WSJ1 corpus, and adapted using MAP estimation with 35 hours of transcribed task-specific training data. The 65k language models are trained on 160 million words of newspaper texts and 132 million words of broadcast news transcriptions. The problem of segmenting the continuous stream of data was investigated using 10 MarketPlace shows. The overall word transcription error of the Nov96 partitioned evaluation test data was 27.1%.