Zooming in and out on phonetic variability and contextual probability

摘要

The phonetic profile of a linguistic unit is known to correlate with its contextual probability, i.e., how likely it is to occur in a given context. In phonetic research, contextual probability is often estimated using simple n-gram models, with bigram probabilities being the most common predictor. While this methodological choice might be seen as merely a technical simplification, it also assumes a highly localized scope of planning in speech production that drives probabilistic phonetic variability. This study assesses this locality assumption by using transformer-based masked language models to estimate probabilities from longer contexts. Results from Taiwan Southern Min and American English demonstrate that larger context windows can lead to improvements in predictions of durational variability. However, comparisons of window sizes and model architectures still suggest a strong role of local contexts.