Acoustical recognition of laryngeal pathology: a comparison of two strategies based on sets of features.

Abstract

The efficiency of sets of acoustical features discriminating pathological voices from control voices is reported. Two strategies were compared. The first (called the 'distance strategy') was built upon a statistical distance of voice features to reference values obtained for a set of healthy (reference) voices. The second strategy (called the 'range strategy') is based on the position inside or outside normal ranges established from a reference population; results based on this strategy were presented in a previous paper. Reference values were calculated from a database of 200 healthy voices distributed into 10-year age groups ranging from 20 to 70. Comparisons were made using a second database of 220 voices, including 65 control, 51 functional dysphonia, 50 with nodules on the vocal folds and 54 recurrent nerve palsy. The phonetic material was compared of 17 French vowels: 11 vowels in a sentence, three isolated vowels and three segments (beginning, middle and end) of the sustained vowel /a/. Four acoustical features were considered for each vowel: the voice fundamental (f0) and the first three formant frequencies. Acoustical features were calculated on an ILS (Interactive Laboratory System) analysis system (workstation). The separation of each pathological group from the control group, using sets of acoustical features, was statistically assessed. From the strategy point of view, results indicated that (i) the fundamental frequency f0 was the best measure to separate normal from pathological voices with the distance strategy; (ii) when the formants were taken, the range strategy performed better in separating the voices. For classification of pathologies, the best separation coefficients were obtained with nodules and the worst with recurrent nerve palsy. Overall, it was seen that the separation between control and pathological voices was most efficient when measured using the distance strategy for f0. The range strategy was useful with formant frequencies.