A speech recognition system with a manual error correction has been developed to produce closed captions in live broadcasting programs such as news programs. Speech recognition errors, however, are not corrected completely because the correctors often miss errors due to lack of care or successive erroneous words. In this paper, we propose a method that detects errors automatically to assist manual correction. A coustic parameters were extracted from both correct and erroneous results of all morphemes produced by speech recognition systems. Templates which can precisely distinguish between errors and correct results, were then constructed by genetic algorithms (GA) and discriminative training. Consequently, theerrors are detected by comparing the acoustic parameters of an unknown recognition result with the templates. Experiments have confirmed that presenting erroneous words by using the proposed method is effective for improving corrector error detection.
