Parametric decomposition of the spectral envelope

In this paper we propose a novel representation of the spectral envelope of speech, using sums of general parametric basis functions (filter bank), including Gaussian, Hann and Nutall. The central frequency and bandwidth of each filter can be adjusted independently for each speech frame as in conventional filter bank analysis. The quality of the representation has been evaluated with respect to a reference spectral envelope obtained from the Harmonic-Stochastic Model [1] (HSM), and also with respect to parameter time stability and separability of different acoustic classes. Numerical results show that the use of the basis functions proposed is an improvement over pure Gaussian models of the spectral envelope.