Enhanced signal classification scheme using a selected information in the ambiguity domain

We present a new time-frequency classification procedure, based on the assumption that in decision problems the redundancy of two-dimensional time-frequency representations should be decreased by investigating only the most interesting parts of the time-frequency representation. Our implementation uses a generic time-frequency representation-the ambiguity function (AF). The classification problem involving several classes can be resolved by generalisation of the ´contrast´ information measure between the mean AFs of all signal classes, or, as we propose, by creating a number of optimal class-to-class comparisons and combine their results as conditional likelihoods. Both classification approaches are demonstrated on a speech signal discrimination problem, for which the latter scheme yields superior results.