Lexically dependent estimation of acoustic information in speech III: Cross-splicing verification of cue weights

In two prior studies (Redmon and Jongman, 2018, 2019, JASA), results of open- and closed-class word recognition tasks on items drawn from a large single-speaker database (Tucker et al., 2018) were used to train a model where acoustic cue weights were optimized to distinguish words in the lexicon, rather than a balanced inventory of phones. From that work, cues were identified that had a greater weight when considering the lexicon as a whole than when studying a symmetric set of contrasts in controlled syllable productions. To verify the causal role of such cues in word recognition, two new cross-splicing versions of the open- and closed-class tasks were run with a subset of items in each prior experiment. In each, an enhancement condition was created by cross-splicing diphones from items in the database with more distinct values on a given cue dimension, and consequently greater model-predicted accuracy. For each such item, a parallel reduction condition was created wherein accuracy was predicted to decrease due to the cross-splicing of a diphone with a more ambiguous value of a given cue. Results will serve to validate the relative role of the different cues in a manner that is external to the model-fitting procedure.