Attempts were made to simulate aspects of the perceptual system by predicting when significant phonetic events are going to occur. The prototype model maintains a grammar of upcoming syllables and periodically computes (1) an instantaneous probability for each possible upcoming segment type using temporal information present at that point, and (2) a probability density function of the onset of prominent acoustic boundaries. Applied to some Japanese data [Port et al., J. Acoust. Soc. Am. (1982)] the model was trained on the speech of five speakers who read a controlled set of Japanese words (like baku, baaku, bakku). It trains by finding the best linear combination of segmental durations to predict (1) locations of future segmental boundaries, as well as (2) syllable‐type possibilities (given the tiny implicit grammar of the experiment). The system was tested on five different speakers reading the same list and was able to anticipate word identity with fair accuracy. Although this system relies on the mora‐timing constraint in order to work (since this regularity makes future events predictable), it does not directly employ the mora in its computation.

This content is only available via PDF.