Neural network-based classification of manner of prevocalic alveolar consonants using temporal formant transitions of vowels in the Buckeye Corpus

• Published in: 음성음운형태론연구 Vol. 30; no. 3; pp. 373 - 399
• Main Author: Soonhyun Hong
• Format: Journal Article
• Language: English
• Published: 한국음운론학회 01.12.2024
• Subjects: 언어학; Buckeye Corpus; temporal formant transitions; prevocalic manner contrasts; neural network
• ISSN: 1226-8690; 2671-616X
• DOI: 10.17959/sppm.2024.30.3.373

The current study aimed to classify the manner of prevocalic alveolar consonants (fricative, lateral, nasal, and stop) in Alveolar-V tokens in the Buckeye Corpus using the temporal formant transitions of the English vowel, vowel category information, F0, gender, vowel duration, the duration of the word containing the token, and the location of the token in the word. A neural network classifier was trained and tested on F1, F2, and/or F3 samples, taken either solely at the vowel onset or taken both at the vowel onset and target. The results show that prevocalic manner could not be manifested properly by the samples of just one or two formants, whether taken solely at the vowel onset or at both the vowel onset and target. However, the classifier trained on all F1, F2, and F3 samples taken at both the vowel onset and target disambiguated prevocalic manner modestly with 67.1% accuracy. The classifier was further trained with additional predictors. Vowel category information considerably improved formant-based classification compared to F0, gender, vowel duration, word duration, and the token’s location in the word. F0 and gender contributed the least. Despite the varying degrees of contribution, these cues collectively enabled the classifier to disambiguate manner contrasts with over 80% accuracy. KCI Citation Count: 0