Tools for Estimating the Perceived Level of Phonetic Reduction

Phonetic reduction is very common in casual speech, where it is associated with several important pragmatic functions. However these phenomena have been little studied. To support investigations and applications, we present tools that automatically estimate the level of perceived phonetic reduction. Trained on annotated dialog data and exploiting HuBert features, these handle American English and Northern Mexican Spanish. For English, word-level predictions correlate up to 0.55 with average human judgments. This is adequate at least for statistical studies of reduction in corpora, as seen in explorations of turn-yielding and prominence-marking behaviors. The tools are open-source and publicly available.