Speech recognition engine for real-time broadcast news captioning

This paper describes a speech recognition engine that progressively outputs the latest available results of words used for real-time closed captioning of Japanese broadcast news. The search engine called a progressive 2-pass decoder practically eliminates the disadvantage of conventional multiple-pass decoders that delay a decision until the end of a sentence. During the first pass of the search the proposed decoder periodically executes the second pass up to that time and detects a part of the final result of words. This method is not theoretically optimal but makes a quick decision with a negligible increase in word errors. In a recognition experiment on Japanese broadcast news, the decoder worked with an average decision delay of 554msec for each word and degraded word accuracy only by 0.22%.