Automated Estimation of Relative Fundamental Frequency

Relative fundamental frequency (RFF), defined as the normalized fundamental frequencies of vowels surrounding voiceless consonants, has been shown to have a characteristic pattern in healthy voices that differs from those with disordered voices (e. g. vocal hyperfunction, Parkinson's disease). However no large-scale clinical study has been performed, mainly because the current estimation protocol requires trained technicians to manually perform this time-consuming task. In this study, we developed a method to automate RFF estimation and tested the algorithm on recordings from 12 healthy participants and 12 participants with Parkinson's disease. The means and variations of RFFs estimated using the automation algorithm were similar to the 'gold standard' estimates developed by two trained technicians. The mean squared error for the automated estimates, when compared to the 'gold standard' RFF estimates, were similar to those estimated manually by an additional trained technician. Future work will focus on improving vocal cycle detection and extending the automation to estimate RFF from instances in running speech.