We introduce a new closed-captioning systembased on automatic speech recognition (ASR) for Japanesebroadcast news programs. In Japanese live closedcaptioning,we have an extreme difficulty in producing captionsmanually because of the lack of a rapid Japanese inputmethod. The ASR technology helps us to make them efficientlyat low cost, but its performance depends largely onacoustic environments. In recognition of noisy speech andinterviews, for example, we do not have enough accuracy forcaptioning. Our newly-developed system employs a "hybrid"ASR approach as the best solution to this problem. Itswitches input speech signals between the original programsound and the rephrased clear speech by a "re- speaker". Itenables the high ASR performance across the entire programsand provides closed-captions with low-latency. In ourtalk, we outline the key features of the system and show thepractical performance in captioning Japanese broadcastnews programs. We also address our experiences and challengesin emergencies, taking up the operation in the GreatEast Japan Earthquake in 2011.
展开▼
