The article concerns techniques for obtaining, representing and comparing voice source signals. Closed-phase formant frequencies and bandwidths were estimated by fitting two linear auto-regressive models to a glottal cycle (the first to the open, the second to the closed phase). The moment of switching from one sub-model to the next was automatically determined by minimizing the overall modelling error. The voice source signal was obtained by inverse filtering speech by means of the closed-phase formants. Its spectrum was represented by a nonlinear zero-memory Volterra model. Two source signals were compared by means of their minimal spectral distance which was obtained by adjusting the nonlinear gain of the Volterra model.