Machine Learning Applied to GRBAS Voice Quality Assessment

Abstract

Voice problems are routinely assessed in hospital voice clinics by speech and language therapists (SLTs) who are highly skilled in making audio-perceptual evaluations of voice quality. The evaluations are often presented numerically in the form of five-dimensional ‘GRBAS’ scores. Computerised voice quality assessment may be carried out using digital signal processing (DSP) techniques which process recorded segments of a patient’s voice to measure certain acoustic features such as periodicity, jitter and shimmer. However, these acoustic features are often not obviously related to GRBAS scores that are widely recognised and understood by clinicians. This paper investigates the use of machine learning (ML) for mapping acoustic feature measurements to more familiar GRBAS scores. The training of the ML algorithms requires accurate and reliable GRBAS assessments of a representative set of voice recordings, together with corresponding acoustic feature measurements. Such ‘reference’ GRBAS assessments were obtained in this work by engaging a number of highly trained SLTs as raters to independently score each voice recording. Clearly, the consistency of the scoring is of interest, and it is possible to measure this consistency and take it into account when computing the reference scores, thus increasing their accuracy and reliability. The properties of well known techniques for the measurement of consistency, such as intra-class correlation (ICC) and the Cohen and Fleiss Kappas, are studied and compared for the purposes of this paper. Two basic ML techniques, i.e. K-nearest neighbour regression and multiple linear regression were evaluated for producing the required GRBAS scores by computer. Both were found to produce reasonable accuracy according to a repeated cross-validation test.