7th International Conference on Spoken Language Processing
September 16-20, 2002
Recently, car navigation systems with a speech interface have been developed. When we communicate with computers through speech recognition, one can not avoid misrecognition and it is difficult to recover from this condition because the interface is only in the initial state of the art. Detection of userís repetition of a misrecognized part can make it easier.
We propose a method to detect partial repetition of misrecognized speech using a word spotting technique based on DTW (dynamic time warping) and N-best hypotheses overlapping measure. We achieved 97.0% detection accuracy for repetitions and 93.5% rejection accuracy for non-repetitions. Next, we tried to improve recognition accuracy using detection. Using the choice of vocabulary setup based on the detection, we achieved improvement in recognition performance in adverse conditions with SNR of 7-13 dB from 29.0% to 35.1% for repaired speech and from 55.5% to 59.3% for non-repetitions. When we employed a rescoring strategy for real time processing, we achieved a 33.5% recognition rate for repaired speech and 57.4% for non-repetitions.
Bibliographic reference. Kakutani, Naoko / Kitaoka, Norihide / Nakagawa, Seiichi (2002): "Detection and recognition of repaired speech on misrecognized utterances for speech input of car navigation system", In ICSLP-2002, 833-836.