7th International Workshop on Models and Analysis of Vocal Emissions for Biomedical Applications (MAVEBA 2011)

Florence, Italy
August 25-27, 2011

Using Acoustic Measures to Predict Automatic Speech Recognition Performance for Dysarthric Speakers

Kinfe T. Mengistu (1), Frank Rudzicz (1), Tiago H. Falk (2)

(1) Department of Computer Science, University of Toronto, Toronto, Canada
(2) Institute National de la Recherche Scientifique, Montreal, Canada

There is growing evidence that clinicians are becoming more receptive to automated computerized tools that assist in treatment decisions and outcomes. Automatic speech recognition (ASR), for example, has had some degree of success as an assistive technology (AT) tool for individuals with mild or moderate dysarthria. Notwithstanding, for a large percentage of individuals with more severe levels of the disorder, ASR has yet to achieve acceptable levels. In this paper, we explore the use of several acoustic measures as correlates of ASR performance for dysarthric speakers. By automatically predicting the potential efficacy of ASR for a particular dysarthric speaker, health care costs and waiting lists may be reduced as may device abandonment rates. Experiments with the "Universal Access" database of dysarthric speech suggest that some of the proposed measures achieve correlations as high as 0.86 with ASR accuracy.

Full Paper (reprinted with permission from Firenze University Press)

Bibliographic reference.  Mengistu, Kinfe T. / Rudzicz, Frank / Falk, Tiago H. (2011): "Using acoustic measures to predict automatic speech recognition performance for dysarthric speakers", In MAVEBA-2011, 75-78.