Repetition induced rhythm as an alternative account to statistical word segmentation: A model and meta-analysis

From a brief exposure to an artificial language, infants and adults can segment out the embedded words in the speech input. The fact that segmentation is successful when the speech contains no prosodic cues to word boundaries has been taken as evidence that humans compute the statistics of syllable transitions to discover the underlying words. However, this explanation cannot account for experiments that show that word segmentation is hindered when the words in the artificial language differ in length, despite the availability of informative statistical properties. In this paper, we offer an alternative theoretical account of these findings. We propose that word segmentation is driven by previously unrecognized rhythmic properties inherent in the artificial language input used in prior studies. In Study 1, we demonstrate that the speech signal used in past studies of word segmentation contains periodicities which are informative as to the length of the embedded words. In Study 2, we present a computational model that uses this rhythmic property to segment speech and to rate possible word candidates. The model not only accounts for cases where learners have been found to segment artificial language streams, but also accounts for cases where they have not, such as when the artificial language stream contains words of variable length. Next, we report on a meta-analysis of the infant statistical word segmentation literature, including segmentation studies with syllable sequences void of natural speech prosody. We found that only sequences made from words that were uniform in length, but not mixed in length, were successfully learned. Taken together, we argue that the perception of rhythm as a fundamental mechanism that is ubiquitous in language and other perceptual domains can likely explain many word-segmentation studies previously attributed to the computation of transitional probabilities.