Abstract:
Previously, a multiscale mechanism [N. Todd and G. Brown, J. Acoust. Soc. Am. 94(A) (1994)] has been shown to be effective in locating significant events in a speech signal and in estimating their relative stress. Two experiments were carried out in order to determine whether major stress points identified by the model correspond to so-called ``p centers'' [Morton et al., Psychol. Rev. 83(5), 405--408 (1976)]. Recordings were made of (1) two subjects reciting the digits ``one'' to ``eight'' and (2) a number of subjects reciting metrical verse in synchrony with a metronome. The speech was recorded on one channel, the metronome on the other. The speech was then extracted without the metronome and fed into the model and the predicted syllabic stress points recovered and compared with the metronome. In the case of the digits the mean error of prediction was -9 ms (N=48). Given that the model has an internal delay of about 50 ms (20 Hz), this is good agreement. The variance was quite high, 1609 ms, but actually within the range of variation typically reported for such a task, e.g., G. Allen [Lang. Speech 15, 72--100 (1972)].