Automatic Classification of Nasals and Semivowels
Tarun Pruthi, Carol Espy-Wilson · 2003
In this paper, we discuss acoustic parameters and a classifier we developed to distinguish between nasals and semivowels. Based on the literature and our own acoustic studies, we use an onset/offset measure to capture the consonantal nature of nasals, and an energy ratio, a low spectral peak measure and a formant density measure to capture the nasal murmur. These acoustic parameters are combined using Support Vector Machine based classifiers. Classification accuracies of 88.6%, 94.9% and 85.0% were obtained for prevocalic, postvocalic and intervocalic sonorant consonants, respectively. The overall classification rate was 92.4% for nasals and 88.1% for semivowels. These results have been obtained for the TIMIT database, which was collected from a large number of speakers and contains substantial coarticulatory effects.