Model variety analysis: determining the need for knowledge-based signal processing

The integration of mathematically formulated signal processing algorithms into signal understanding systems is addressed. If control parameter adjustment is found to be necessary, a signal processing algorithm is said to have a model variety problem with respect to the class of input signals in the application domain. The necessity for such parameter adjustments implies that knowledge-based techniques must be used to adjust the signal processing parameters. A systematic framework is presented for analytically determining whether a given signal processing algorithm has a model variety problem with respect to a particular application domain. The framework is illustrated for the short-time Fourier transform algorithm as applied to the problem of non-speech sound analysis.>