Speech segmentation (summary)


Background As infants begin to learn the language spoken around them, they face the problem of 'speech segmentation'. That is, infants hear speech in a continuous stream without any pauses between words, yet, in order to acquire language, they must learn how to separate continuous speech into individual words.

What was the aim of the study? The authors of this study designed and tested a computational model of speech segmentation.

Why was the paper written? While other computational models of speech segmentation exist, the authors suggest a set of criteria that the most successful models should share. Models of speech segmentation should be able to identify individual words, plausibly reflect the cognitive processing abilities of an infant, have a limited number of external components and reflect what researchers know about how infants learn to speak. The authors developed a new model that could more completely meet these criteria than older models.

What did the authors do? The authors built a computational model of speech segmentation. They focussed on using boundaries in speech, such as pauses between words and changes in speakers, and word frequency to segment speech. The model segmented continuous speech based on whether any components of the speech were familiar. For example, if the model encountered a word in isolation, such as 'car', that word could be identified when it occurred in longer strings of speech, such as 'look it car'. If another phrase contained the phrase 'look it', then the model could recognize 'look it' as an independent item.

The authors were interested in whether the model would be able to identify single words, what words were identified most easily and how words were discovered over time.

What did they find? The model was able to successfully identify many real words in the data, which included speech that adults directed to six different children aged two-and-a-half years and younger. The model was able to identify single words quickly, including the children's names. The authors found that the words that the model segmented overlapped with the kinds of words that are useful for the acquisition of grammar.

Conclusions: The model designed by these authors was just as successful as older models of speech segmentation, but it fulfilled the criteria of successful models better than others. The model demonstrated that a child's name may have an important role in speech segmentation, as children may learn to recognize their name early in development and use this knowledge to segment the speech they hear surrounding their names. Importantly, the words that were useful for segmenting speech were also useful for grammar acquisition, which suggests that the same words may form an important basis for both skills.

Things you may want to look into:

Developing language in a developing body (summary)

Working memory in children's language comprehension (summary)

Added to site August 2013