Predicting Normal and Pathological Voice using a Cepstral Based Acoustic Index in Sustained Vowels versus Connected Speech

Objectives: In recent years, the use of cepstral measures for acoustic evaluation of voice has increased. The objective of this study is to evaluate the diagnostic value of spectral/ cepstral measures to differentiate dysphonia from normal voice and to determine what type of voice sample (sustained vowel /a/ or connected speech) is the most sensitive in differentiating normal and pathological voice. Methods:Two hundred and eighty-eight individuals (99 men, 189 women) from 214 dysphonia patients and 74 normal speakers recorded connected speech and a sustained vowel /a/. One laryngologist and two speech-language pathologists performed visual and auditory-perceptual rating of voice samples in terms of the degree of dysphonia/normality. Recorded voices were analyzed with two spectral/cepstral measures.The cutoff value for positivity that has the highest specificity for discriminating between normal and dysphonia voices was determined based on receiver operating characteristic (ROC) analyses. Results: Measures of cepstral peak prominence (CPP) and ratio of low- to high-frequency spectral energies (UH ratio) were significantly different between groups in both speech conditions. ROC analysis demonstrated CPP had high sensitivity and specificity for the classification of dysphonia versus controls in the both speech conditions (area under curve [AUC] = .815 in vowel, AUC= .91 in connected speech); and CPP, in particular, showed higher discrimination accuracy. Conclusion: CPP is a good predictable acoustic measure to detect dysphonic speakers in both vowel prolongation and connected speech from normal voice.Therefore, this study suggested cepstral-based acoustic measures should be included for clinical evaluation of dysphonia.