Automating quasi-stationary speech signal segmentation in sustained vowels: application in the acoustic analysis of Parkinson’s disease

Acoustic analysis of sustained vowels is typically used to quantify perturbations in fundamental frequency (F0), amplitude, and deviations from periodicity, and associate these with clinical outcomes of interest. Computational and practical constraints suggest that 2-3 seconds are often sufficient to acoustically characterize a sustained vowel phonation. The question then is how to best determine a short quasi-stationary segment from a typical 20-30 seconds speech recording. We computed the F0 contour in 10 millisecond epochs using SWIPE, a state-of-the-art F0 estimation algorithm, which we had previously demonstrated is very competitive in F0 estimation for sustained /a/ vowels. Subsequently, we determined the two second signal segment that exhibits the smallest mean absolute successive F0 difference. We tested the segmentation algorithm on 100 randomly selected sustained vowel /a/ phonations from the Parkinson’s Voice Initiative, where we had hand-labeled the quasi-stationary segments. We found the algorithm correctly identified the quasi-stationary segments in all cases, thus demonstrating it can be deployed at large scale studies automating further processing of sustained vowels. We also demonstrated that this pre-processing step can have a major influence in the acoustic characterization of the phonations.