On separating voiced-speech into its components

We present a new strategy and several algorithms to separate voiced-speech into its natural components. Each component is a generalized spectral line whose instantaneous frequency and amplitude continually evolves in time. The decomposition is achieved by a set of demodulators followed by an array of interconnected feed-forward/feed-back filters. Further analysis of the separated components shows that the instantaneous frequencies of some of them (especially those which are located close to a narrow formant) can significantly deviate from their harmonic location, for tens of milliseconds. The envelopes of some components also exhibit interesting amplitude modulations. We conjecture that the group-delay compensated feed-back filter (RISC) structure we propose is similar to the bandpass filters implemented in the mammalian cochlea