The article discusses developing a system for close-captioning and automatic information extraction for Japanese broadcast news speech. The article presents a large-vocabulary, continuous-speech recognition system for Japanese broadcast news speech transcription. It is part of a joint research project with NHK broadcasting whose goal is the closed-captioning of TV programs. While some of these problems investigated are Japanese-specific, others are language independent. A work-frequency list was derived for the news manuscripts, and the 20,000 most frequently used words were selected as vocabulary words. The feature vector consisted of 16 cepstral coefficients, normalized logarithmic power, and their delta features. Cepstral coefficients were normalized by the cepstral mean subtraction method. The acoustic models were gender-dependent shared-state triphone hidden Markov models and were designed using tree-based clustering.