An application of continuous speech recognition to superimposing captions

Abstract

We have developed experimental equipment which recognizes continuous speech uttered at a "Bunsetsu" unit by specified speakers such as announcers, and superimposes the recognition results as TV captions. The speech recognizer is implemented on micro parallel processors. In a recognition experiment on announcement speech for Sumo TV programs, all sentences including wrestlers' names and winning tricks are correctly recognized and superimposed. Speech recognition of real hall announcement is now under consideration.