Investigating the temporal distribution of words in child-centered audio using the burstiness metric

As the use of naturalistic data of children's language experiences increases, the temporal dynamics of the environment becomes a more apparent feature of naturalistic language. By investigating the temporal dynamics of the language environment, we can both more accurately describe the experiences of young children and connect the temporal dynamics of child-centered speech to existing experimental work showing that the temporal distribution of training has measurable effects on learning. To describe temporal patterns in children's experiences, the first step is to develop measures that quantify temporal patterns in naturalistic speech. The current work examines one such measure, the Burstiness metric, and investigates individual words’ burstiness and its relationship to frequency, its behavior across different lexical classes and timescales, and whether it can be used to quantify experimental constructs of temporal orders. Using datasets of both daylong audio recordings and shortform recordings of child-centered speech, our findings show that while related, word burstiness is not an index of word frequency. Additionally, our results show that burstiness varies across words of different lexical classes and that burstiness hinges on timescale, such that words that are bursty at longer timescales are not necessarily the same ones that tend to be similarly bursty at shorter timescales. Our findings emphasize that the irregular structure of speech consequently affects how burstiness may be interpreted at different timescales and highlights how the Burstiness metric does not neatly map on to our verbal theories of temporal patterns. We discuss implications for how this measure may be used for child-centered audio and potential pitfalls.