Optimal estimation of subband speech from nonuniform non-recurrent signal-driven sparse samples

Speech signals are comprised of auditory objects that are localized in time, but can appear anywhere in the record. We introduce a strategy for non-recurrent irregular signal-driven sampling and subsequent maximum likelihood interpolation of speech subbands that achieves object constancy—the representation of an auditory object is precisely locked to the timing of its features, but is otherwise constant. Moreover, the reconstruction fidelity can be traded flexibly for sampling rate, over a broad range of signal-to-noise ratios and application requirements. In an experiment with wideband speech, we find a regime in the rate/distortion curve that has almost perfect reconstruction at a rate substantially lower than the respective Nyquist rate.