This paper presents a comparative study of model versus non-model based classifiers, which are systematically evaluated on two standard tasks in speech processing. Model based approaches are represented by a Gaussian classifier, non-model based approaches by a multi-layer perceptron. The performance of both classifiers is evaluated with respect to (1) a acoustics-to-phoneme class recognition task, and (2) a limited grapheme-to-allophone conversion task. The Gaussian classifier introduced in this study differs from conventional designs in that the density parameters, i. e. the mean and covariance values, are established by way of optimization. An iterative, gradient-based, algorithm is employed to determine the parameter values that will result in minimal mis-classification. The results of the evaluations presented in this study indicate that the Gaussian classifier performs at least as well as the MLP-based classifier, with respect to both discrimination tasks.
Bibliographic reference. Brauer, Peter / Hedelin, Per / Huber, Dieter / Knagenhjelm, Petter / Molno, Johan (1991): "Model or non-model based classifiers", In EUROSPEECH-1991, 1027-1030.