5th International Conference on Spoken Language Processing
Although the performance of state-of-the-art automatic speech recognition systems on the challenging task of broadcast news transcription has improved considerably in recent years, many of the systems operate in 130-300 times real-time. Many applications of automatic transcription of broadcast news, eg. closed-caption subtitles for television broadcasts, require real-time operation. This paper describes a connectionist-HMM system for broadcast news transcription, and the modifications to this system necessary for real-time operation. We show that real-time operation is possible with a relative increase in word error rate of about 12%.
Bibliographic reference. Cook, Gary / Robinson, Tony / Christie, James (1998): "Real-time recognition of broadcast news", In ICSLP-1998, paper 0065.