Information-theoretic hypothesis generation of relative cue weighting for the voicing contrast

To learn the voicing contrast, children must identify which of the available perceptual cues are helpful in different contexts. Using Standard American English (SAE) as a case study, we generated hypotheses of which cues are the most informative for different contexts, such as onsets vs. codas. More specifically, we classified SAE obstruents as voiced vs. voiceless using decision trees trained and tested on TIMIT. We validated the feature importances of different contexts against the findings of previous perceptual studies and we gleaned more specific hypotheses to help design future experiments on children's acquisition of the voicing contrast.