ISCA Archive Interspeech 2024
ISCA Archive Interspeech 2024

Phonological Feature Detection for US English using the Phonet Library

Harsha Veena Tadavarthy, Austin Jones, Margaret E. L. Renwick

This paper details adaptation of the Phonet Library, a speech technology that uses distinctive features to calculate posterior probabilities for phonological classes, by training it on US English. Phonet's posterior probabilities provide a statistical basis for understanding patterned variability in speech, thus bridging the gap between acoustic data and phonological structures. We train Phonet on an English corpus, and we investigate both its precision in phoneme recognition at designated timestamps and the relationship of its distinctive feature probabilities to linguistic expectations for a subset of vowels and consonants. Phonet shows accuracy of over 90\% in predicting English phonological classes. Our analysis demonstrates Phonet's robustness in capturing basic relationships between theoretical natural classes of sounds in English, highlighting its utility in the broader context of speech analysis and phonetic research.