Automatic discovery of acoustic measurements for phonetic classification
Michael Phillips · The Journal of the Acoustical Society of America · 1988
One approach to phonetic classification is to rely on the use of well-motivated acoustic measurements such as formant frequencies. However, the extraction and the use of these measurements are often heuristically based, thus resulting in fragile classification performance. Alternatively, one can rely completely on learning procedures to automatically discover the phonetic regularities. But the requirement on computation and training data may become prohibitive. This paper describes an approach in which the determination of the acoustic measurements is formulated as a constrained search problem. In the present investigation, the constraints are provided through a set of generic measurement primitives based on knowledge of acoustic phonetics. These primitives, such as spectral moments or energy changes, have free parameters that control the specifics of the measurements. Primitives may be combined to form more complex measurements according to a set of constraints. The set of primitives, along with their free parameters and the combination constraints, specify a space of allowable measurements. This space can be searched using a measure of phonetic discrimination performance on a large body of training data to obtain an optimal set of measurements for a particular task. Results of classification experiments using these automatically generated measurements will be presented for various phonetic discrimination tasks. [Work supported by DARPA-ISTO under contract N00039-85-C-0254.]