Automatic Detection of Hypernasality in Cleft Palate Speech Using Phase Feature

The detection of hypernasality, a speech disorder common in children with cleft palate, typically involves analyzing vowels within speech stimuli. An essential acoustic cue for hypernasality detection is the presence of a nasal peak near the first formant
$(F_{1})$
. Traditionally, a two-step process is employed, where vowel selection from stimuli is initially performed through manual annotation, followed by automatic feature extraction and classification. This work introduces a novel method for hypernasality detection, automating the entire process, including vowel region selection, feature extraction, and classification. The method leverages vowel onset points for the automatic identification of the vowel region. Subsequently, a modified group delay cepstral coefficient feature is extracted from this automatically selected vowel region. The feature is extracted from modified group delay spectrum, offering enhanced resolution to distinguish closely spaced nasal peaks and the
$F_{1}$
, thus improving the capture of nasality evidence. The proposed method demonstrates promising results when applied to hypernasality detection using a support vector machine classifier with the accuracy of 83.50% for /a/ vowel and 93.21% for /i/ vowel.