Use of Prosodic Features in Speech Recognition

Two methods were proposed for the use of prosodic features in speech recognition: one to detect major syntactic (phrase) boundaries as the initial phase of speech recognition, and the other to check the feasibility of the results of ordinary recognition process from the viewpoint of prosodic features. In the rst method, fundamental frequency contours were assumed as waveforms as functions of time and were low-pass ltered to suppress accent components in the contours. Then the derivative of ltered contour was used to detect phrase boundaries. An experiment was conducted on the ATR continuous speech database, showing that the method managed to detect about 77% of manually detectable phrase boundaries. The second method is based on generating fundamental frequency contours for recognition candidates using a speech synthesis scheme and comparing them with the observed contour. The candidate giving the best matched contour to the observed contour should be the nal recognition result. The method was shown to be valid in detecting recognition errors accompanied by changes in accent types or/and in syntactic boundaries. The method was then evaluated in its performance for the detection of phrase boundaries. Allowing 1-mora discrepancies, the detection rate reached 92% for the ATR database, which was further improved to 97% by a simple speaker adaptation method.