A maximum peakiness criterion for deconvolving speech waveforms

Linear predictive coding of speech has traditionally used a least mean square (LMS) criterion for determining the optimal filter parameters. Auto-regressive (AR) models using LMS are in wide-spread use for low bit rate speech compression. These models are particularly susceptible to moderate amounts of added noise and distortion. This paper proposes a new model which does not use the least-mean-square criterion but rather uses the "peakiness" of the deconvolved waveform to determine the filter parameters. Since voiced output speech is reproduced with a maximally "peaked" waveform, viz, an impulse, this criterion is naturally suited for the characteristics of the synthesis process. Preliminary analysis also shows the feasability of implementing the proposed algorithm in microprocessor based hardware.