Computerised objective measurement of strain in voiced speech

Area of Science:

• Otolaryngology
• Speech Science
• Biomedical Engineering

Background:

• Healthcare professionals require objective voice quality assessment for patients with voice disorders.
• The subjective GRBAS (Grade, Roughness, Breathiness, Asthenia, Strain) scale, used by Speech and Language Therapists (SLTs), is effective but costly.
• Objective measurement of voice properties, particularly 'Strain', is needed to complement subjective assessments.

Purpose of the Study:

• To develop and compare computerised objective measurement models for voice 'Strain'.
• To evaluate the accuracy of K Nearest Neighbor Regression (KNNR) and Multiple Linear Regression (MLR) models against SLT assessments.
• To identify optimal feature subsets for objective voice strain prediction.

Main Methods:

• Two regression models, KNNR and MLR, were employed for objective voice 'Strain' measurement.
• Models were trained and tested using various feature subsets and datasets.
• Performance was evaluated by comparing model predictions against the mean scores of five experienced SLTs on 102 voice samples.

Main Results:

• The study identified the best feature subsets for objectively predicting voice 'Strain'.
• Computerised measurements achieved a lower Normalized Root Mean Square Error (NRMSE) compared to individual SLT scores.
• MLR and KNNR models yielded NRMSEs of 14.6% and 15.1%, respectively, using optimal feature subsets.

Conclusions:

• Objective voice 'Strain' measurement using MLR and KNNR models is a viable and accurate alternative to subjective SLT assessments.
• Computerised methods offer a cost-effective and consistent approach to voice quality evaluation.
• These findings support the integration of objective computational tools in clinical voice analysis.