Testing an automated subtitle-generation system based on speech-recognition technology at broadcast stations in seven prefectures

摘要

We have conducted experimental tests of a cloud-based artificial intelligence (AI)-enabled speech-recognition system that automatically generates and displays subtitles for television programs from regional broadcasting stations. Subtitles are a broadcast service in which the spoken-word elements of television programs, including narrations, dialogue between performers, and other speech content, are displayed in the form of on-screen text as a viewing aid for hearing-impaired viewers, elderly viewers, and other viewers for whom spoken-word content is difficult to understand. For live news broadcasts, subtitles are typically generated by a combination of two methods. In the first method, multiple human operators listen to, and manually transcribe, the words spoken by news presenters. In the second method, automated speech-recognition algorithms are used to convert spoken words into transcription text. These are then checked by human operators to correct inaccuracies due to speech-recognition errors. Ideally, subtitle broadcasts should be available not only for news broadcasts but for a wide range of programs, including those offered by regional broadcasting stations. In practice, however, it is often difficult to secure the manpower and equipment required to generate subtitles, which means subtitle broadcasts are not available for some subset of programs.