Frequency-time methods in spectral analysis [speech]

The success of spectral analysis in dealing with stationary processes makes it a natural candidate for extension to dealing with nonstationary processes. Spectra reveal insight into the structure of a process and aid model fitting, and they may be estimated by simple methods. With this in mind, it is desirable that any generalisations to nonstationary processes should preserve as many aspects of this as possible. There are many candidates for spectra and the authors only note a few, highlighting those that use both frequency and time together yielding the concept of a spectrum that changes with time. The authors attempt to: highlight the basic considerations in using both time and frequency to characterise signals, note some desirable properties of frequency-time spectra, define the evolutionary spectrum and the Wigner-Ville spectrum, comment on estimation methods for the above spectra, together with other (empirical) procedures, note a (recent) modification to the moving window approach due to Kodera, and present some results to show how the above methods compare