Representations of sound that are insensitive to spectral filtering and parametrization procedures

This paper describes representations of time-dependent signals that are invariant under any invertible signal distortion. Such a representation can be created by rescaling the signal in a nonlinear dynamic manner that is determined by recently encountered signal levels. Information that is encoded in such representations will be faithfully communicated in the presence of severe signal distortions, which may originate in the transmitter, receiver, or the channel between them. As in speech communication, the receiver is "blind" and need not characterize the form of the signal distortion, which remains unknown. The method is applied to analytical examples, acoustic waveforms of human speech, and the short-term Fourier spectra of a bird song. The results suggest that the rescaled representation of a sound is insensitive to the way its spectra have been filtered and parametrized, as long as those processes do not obliterate the differences between the various spectra in the sound. Finally, the possible "speaker" independence of these representations is explored in the context of a simple linear prediction model of vocal tracts with a single degree of freedom.