Related Experiment Videos
Novel approach to acoustical voice analysis using artificial neural networks
R Schönweiler1, M Hess, P Wübbelt
1Department of Communication Disorders, Center of Ophthalmology, Otorhinolaryngology and Communication Disorders, Hannover Medical School, Germany. schoenweiler.rainer@mh-hannover.de
Journal of the Association for Research in Otolaryngology : JARO
|September 8, 2001
Summary
This study explored using artificial neural networks to analyze acoustic voice parameters and correlate them with perceptual voice quality ratings. Feedforward networks showed promising accuracy, suggesting potential for automated voice disorder assessment.
Area of Science:
- Speech science
- Biomedical engineering
- Machine learning
Background:
- Perceptual rating scales are standard for voice quality assessment but are subjective.
- Acoustic measures offer objective voice analysis, potentially correlating with subjective perception.
- Multivariate statistics and artificial neural networks (ANNs) can identify complex patterns in acoustic data.
Purpose of the Study:
- To investigate if ANNs and multivariate statistics can identify acoustic voice parameter patterns corresponding to perceptual voice quality ratings (RBH index).
- To develop a classification system for acoustic voice analysis based on ANN performance.
Main Methods:
- A multicenter study rated voice samples from 117 individuals using the RBH index (roughness, breathiness, hoarseness).
- Acoustic features were extracted and analyzed using multivariate regression trees and ANNs (LVQ and RProp algorithms).
- Feedforward networks were selected for developing a classification system.
Main Results:
- Feedforward networks achieved classification accuracies of 65-85%.
- A classification system using 50 feedforward networks showed 40% accuracy matching a priori RBH values and 65% matching at least one domain.
- These accuracies align with other ANN applications in biology and clinical medicine.
Conclusions:
- Feedforward neural networks demonstrate significant potential for objective acoustic voice analysis.
- The developed classification system shows promise for automated voice disorder assessment.
- Further research into ANNs for acoustic voice analysis is encouraged.