Description:

How context can help in learning sounds from naturalistic speech

Infants learn the sound categories of their language and adults successfully process the sounds they hear, even though sound categories often overlap in their acoustics. Most researchers agree that listeners use context (e.g. who the speaker was, what the neighboring sounds were, etc.) to help disambiguate overlapping categories, and have put forth a number of theories about how contextual information could be used. However, for the most part these theories have been developed by studying simplified speech (synthetic or well-enunciated, controlled lab speech), so it is unclear to what extent these ideas extend to naturalistic speech. Here, I ask how contextual information could be helpful for processing and learning from naturalistic speech of the type that listeners actually hear. I implement two main ways of using context and test their efficacy in separating overlapping categories on naturalistic speech, focusing on the test case of Japanese vowel length. Our results show that well-established results from lab speech do not necessarily generalize to naturalistic speech, and lead to a new proposal for how infants could learn the sounds of their language. Overall, our results reveal the importance of studying infants’ naturalistic input and highlight the value of tools that allow us to do so.