RNN Classification of English Vowels: Nasalized or Not
- Ling Liu (University of Colorado Boulder)
- Mans Hulden (University of Colorado Boulder)
- Rebecca Scarborough (University of Colorado Boulder)
Abstract
Vowel nasality is perceived and used by English listeners though it is not phonemic. Feature-based classifiers have been built to evaluate what features are useful for nasality perception and measurement. These classifiers require heavy high-level feature engineering with most features discrete and measured at discrete points. Recurrent neural networks can take advantage of sequential information, and has the advantage of freeing us from high-level feature engineering and potentially being stronger simulation models with a holistic view. Therefore, we constructed two types of RNN classifiers (vanilla RNN and LSTM) with MFCCs of the vowel as input to predict whether the vowel is nasalized or not. The LSTM model achieved the best performance, and supports the phonetic claim about the degree of coarticulatory nasality and the use of MFCCs for automatic speech recognition.
Keywords: vowel nasalization, classification, recurrent neural network, perception
How to Cite:
Liu, L., Hulden, M. & Scarborough, R., (2019) “RNN Classification of English Vowels: Nasalized or Not”, Society for Computation in Linguistics 2(1), 318-321. doi: https://doi.org/10.7275/cnem-ab98
Downloads:
Download PDF