Files
Abstract
This research presents re-training of Phonet Library, a speech technology that calculates posterior probabilities for phonological classes by leveraging distinctive features, on an American corpus. We re-trained the model on 49 phonemes classified by 24 distinctive features + pause (silence). We call the resulting model Phonet_English. It considers both the acoustic features and the phonetic features to estimate the posterior probabilities for a given audio signal. This statistical approach helps us understand patterned variability in speech. Phonet_English exhibits an impressive range of accuracies for phonological class recognition, with the lowest accuracy value of 80.7% and the highest accuracy value of 96.3%. This thesis also delves into the model’s phoneme recognition accuracy and examines how its distinctive feature probabilities align with linguistic expectations for selected vowels and consonants. Our results showcase that Phonet_English is successful in capturing fundamental relationships between theoretical natural classes of sounds and their realization in English, making it highly useful in speech analysis and phonetic research.