A data-driven scheme for the approximated computing of alias-free generalized discrete time-frequency distributions

The time-frequency distribution (TFD) based on Cohen's (see Time-Frequency Analysis. Signal Processing Series. Prentice Hall PTR, New Jersey, 1995) class has significant potential for the analysis of a number of non-stationary signals. One of the discrete formulations is the recently introduced alias-free generalized discrete-time TFD (AF-GDTFD). The spectral decomposition of the kernel allows the computation of AF-GDTFD as a weighted sum of spectrograms. The partial sum has been shown to offer a vehicle to trade-off between exactness and computational load. This paper proposes a scheme which exploits local approximations by adapting dynamically the accuracy of spectrograms to the eigenvalue magnitudes. The approach employs the wavelet packet transform followed by a block-recursive Fourier transform and a compensation network. Adaptive selection of subbands for further processing reduces substantially the computational cost while still preserving an acceptable quality. The approach is attractive in terms of VLSI aspects due to the modular structure, local connections and stream processing.