Detection and recognition of Repaired Speech on Misrecognized Utterances for Speech Input of Car Navigation System

Abstract

Recently, car navigation systems with the interface based on speech recognition have been realized. When we communicate with computers through speech interface, we cannot avoid misrecognition. However, it is difficult to recover it because of the immaturity of the interface. Detection of user's repetition of misrecognized part can make it easier.
We propose a method to detect the partial repetition of misrecognized speech using a word spotting technique based on DP matching and N-best hypotheses overlapping measure. We achieved 97.0% of detection accuracy for the repetition and 93.5% of the rejection accuracy for others.
Next, we tried to improve the recognition accuracy using the detection. Using the choice of vocablary setup based on the detection, we achieved improvement of the recognition performance from 29.0% to 35.1% for repaired speech and from 55.5% to 59.9% for others. When we employed restoring strategy for real time processing, we achieved 33.5% of recognition rate for repaired speech and 58.5% for others.