Application of classifier-optimal time-frequency distributions to speech analysis

Discrete operator theory maps each discrete time signal to a multitude of time-frequency distributions, each uniquely specified by a kernel function. This kernel function selects some details to emphasize and other details to smooth. Traditionally, kernels are chosen to impart specific properties to the resulting distributions, such as satisfying the marginals or reducing cross-terms. Given a labeled set of data from several classes, we seek to generate a kernel function that emphasizes classification relevant details present in the distribution. In this paper, we extend our previous work on class dependent time-frequency distributions. Previously, the discriminant function did not consider the within-class to between-class variance of coefficients, and was vulnerable to choosing very “noisy” features