Accurate Representation of Time-Varying Signals Using Mixed Transforms with Applications to Speech

An approach that uses mixed partial sets of basis functions is proposed for time-varying signal representation. Basis functions from two transforms are used, namely, the discrete Fourier transform (DFT) and the Walsh-Hadamard transform (WHT). An iterative algorithm is developed to implement the mixed representation. A finite number of DFT components, either fixed or variable, is used to represent the selective (narrowband) dominant spectrum of the signal on a frame-by-frame basis. The whitened (broadband) residual signal is represented using a small number of Walsh components. The mixed partial sets are applied to speech representation. Preliminary results and informal listenings using this model produced good quality speech.