Automatic detection of vowel nasalization using knowledge‐based acoustic parameters

Researchers in the past have suggested several acoustic correlates of nasalization including extra pole‐zero pairs near the first formant (F1), a reduction in F1 amplitude, and an increase in F1 bandwidth. Even though these correlates have been known for a long time, considerable work is still needed to automate the extraction of acoustic parameters (APs) for nasalization. This work looked at 37 different APs which were pared down to 8 APs based on F statistic obtained by ANOVA. In preliminary experiments, an accuracy of 69.79% has been obtained for the task of discriminating between oral and nasalized vowels on the TIMIT database using a support vector Machine (SVM)‐based classifier. The classification was done on a frame basis, and a segment was declared nasalized if more than 30% of the frames were found to be nasalized. Note that all vowels adjacent to nasal consonants were assumed to be nasalized. Thus, the accuracy may actually be higher since some vowels before nasal consonants may not be nasalized. Further, these results were obtained by using a linear kernel in SVMs. We hope the results would improve when a radial basis function kernel is used. [Work supported by Honda and NSF Grant BCS0236707.]