Self‐organizing map for the classification of normal and disordered voices

The classification of voice disorders has been hampered by the questionable reliability of perceptual scaling methods and by the generally poor correlations between perceptual and acoustic data. This report uses a self‐organizing map to classify normal and disordered voices. The self‐organizing feature map (SOFM) is essentially a two‐dimensional representational mapping over an array of processing units of the multidimensional nonlinear regularities inherent in the input data space. The input data were five acoustic measures obtained from a MDVP [Multi‐Dimensional Voice Protocol, Kay Elemetrics Corp.] analysis of sustained phonations included in the Voice Disorders Database version 1.03, compiled by the Voice and Speech Laboratory, Massachusetts Eye and Ear Infirmary. The measures included: jitter percent, fundamental frequency variation, shimmer percent, noise‐to‐harmonics ratio, and amplitude of the dominant cepstral rhamonic. The SOFM was trained with three voice samples: normal, hyperfunction, and anterior–posterior squeezing. The SOFM also was used to classify voices obtained from individuals with neurogenic speech disorders (dysarthrias). [Work supported by NIH.]