Directly test the proposed mechanism linking augmentation to phonological errors

Test directly whether acoustic signaling properties, low training exposure, and their interaction explain the phonological error patterns produced by SpecAugment, and determine why Tag-Based Replacement reduces palatalization errors without changing exposure.

Background

The paper proposes that SpecAugment disproportionately harms categories that are both infrequent in training and signaled by a single acoustic property. Tone and palatalization fit this description, whereas stress does not, and the observed error patterns are consistent with the account.

The authors acknowledge that the relevant causal links have not been directly tested. In particular, signaling is inferred from phonetic knowledge, low exposure and single-property signaling coincide for the studied tone and palatalization categories, and the exposure-independent reduction in palatalization errors from Tag-Based Replacement remains unexplained.