Third International Conference on Spoken Language Processing (ICSLP 94)
This paper proposes a new method for estimating speaking rates of Japanese using a speech recognition technique. Taking a fact that a Japanese syllable generally contains one and only one vowel into consideration, the method estimates speaking rates based on a vowel recognition technique. The vowel recognition is performed using the LPC cepstral distance between the input and pre-stored vowel templates which have been trained for speaker-independent vowel recognition. The new method has been applied for two definitions of speaking rate: one is the number of moras in a certain time interval, and the other is the inverse of an average period of successive vowels. Experiments showed that the estimation errors of speaking rate with those two definitions are 3.0% and 8.7%, respectively.
Bibliographic reference. Ando, Akio / Miyasaka, Eiichi (1994): "A new method for estimating Japanese speech rate", In ICSLP-1994, 731-734.