Automatic speech and singing classification in ambulatory recordings for normal and disordered voices

• DOI: 10.1121/1.5115804
• Publication Date: 2019-07-11T15:26:22Z
• Authors: Andrew J. Ortiz; Laura E. Toles; Katherine L. Marks; Silvia Capobianco; Daryush D. Mehta; Robert E. Hillman; Jarrad H. Van Stan
• Keywords: Adult; 03 medical and health sciences; Voice Disorders; 0302 clinical medicine; Voice; Humans; Monitoring, Ambulatory; Singing; Speech; Female; Vocal Cords; 0305 other medical science

Ambulatory voice monitoring is a promising tool for investigating phonotraumatic vocal hyperfunction (PVH), associated with the development of vocal fold lesions. Since many patients with PVH are professional vocalists, a classifier was developed to better understand phonatory mechanisms during speech and singing. Twenty singers with PVH and 20 matched healthy controls were monitored with a neck-surface accelerometer–based ambulatory voice monitor. An expert-labeled ground truth data set was used to train a logistic regression on 15 subject-pairs with fundamental frequency and autocorrelation peak amplitude as input features. Overall classification accuracy of 94.2% was achieved on the held-out test set.