Spectral estimation by damped sinusoidal modelling and singular value decomposition [speech recognition]

Looks at how to estimate spectra of decaying signals. It is interesting to note that there are again, two different methods. The first relies on forming a prediction filter (like LPC) and the second uses time-varying narrowband filters. Since the data is decaying, a LPC lattice structure is not appropriate, but signal-to-noise enhancement can be accomplished through the use of the singular value decomposition (SVD). The decaying signals lead to a need for unstable narrowband filters