Abstract:
An analysis of spoken language in lecture style was carried out, where two kinds of pauses, acoustic pauses and perceptual pauses, were defined and detected from the same speech material. While the former were defined as speech segments with lower power than a given threshold and detected semi-automatically, the latter were defined as speech segments during which humans perceive a ``break'' in the utterances and obtained through perceptual experiments. After detecting both the pauses, the probability of an acoustic pause being also a perceptual pause was investigated separately for each of the speakers. As a result, it was found that the probability for the acoustic pauses of a given length was largely decreased when the speaking rate was lower. This implies that, especially in the case of pause-based speech processing, the pauses should be processed considering the speaking rate around them. Furthermore, the investigation of interjections and filled pauses around the perceptual pauses showed that ``e[e],'' ``e[e]to,'' and ``de'' caused the perceptual pauses with quite high probability.