A practical method of estimating the time-varying degree of vowel nasalization from acoustic features

This paper presents a simple and easy-to-use method of creating a time-varying signal of the degree of nasalization in vowels, generated from acoustic features measured in oral and nasalized vowel contexts. The method is presented for separate models constructed using two sets of acoustic features: (1) an uninformed set of 13 Mel-frequency cepstral coefficients (MFCCs) and (2) a combination of the 13 MFCCs and a phonetically informed set of 20 acoustic features of vowel nasality derived from previous research. Both models are compared against two traditional approaches to estimating vowel nasalization from acoustics: A1-P0 and A1-P1, as well as their formant-compensated counterparts. Data include productions from six speakers of different language backgrounds, producing 11 different qualities within the vowel quadrilateral. The results generated from each of the methods are compared against nasometric measurements, representing an objective "ground truth" of the degree of nasalization. The results suggest that the proposed method is more robust than conventional acoustic approaches, generating signals which correlate strongly with nasometric measures across all vowel qualities and all speakers and accurately approximate the time-varying change in the degree of nasalization. Finally, an experimental example is provided to help researchers implement the method in their own study designs.