Sampling children's spontaneous speech: how much is enough?

There has been relatively little discussion in the field of child language acquisition about how best to sample from children's spontaneous speech, particularly with regard to quantitative issues. Here we provide quantitative information designed to help researchers make decisions about how best to sample children's speech for particular research questions (and/or how confident to be in existing analyses). We report theoretical analyses in which the major parameters are: (1) the frequency with which a phenomenon occurs in the real world, and (2) the temporal density with which a researcher samples the child's speech. We look at the influence of these two parameters in using spontaneous speech samples to estimate such things as: (a) the percentage of the real phenomenon actually captured, (b) the probability of capturing at least one target in any given sample, (c) the confidence we can have in estimating the frequency of occurrence of a target from a given sample, and (d) the estimated age of emergence of a target structure. In addition, we also report two empirical analyses of relatively infrequent child language phenomena, in which we sample in different ways from a relatively dense corpus (two children aged 2;0 to 3;0) and compare the different results obtained. Implications of these results for various issues in the study of child language acquisition are discussed.