Speech Coding Based on Time-Varying Bit Allocations to Excitation and LPC Parameters

In coding systems based on the excitation- and -LPC model for speech, parameters of the two parts of the model are separately quantized, with levels of fidelity that are constrained by the total bit rate available to the system. For reasons of simplicity and for the lack of a good design criterion, the bit rates allocated to the excitation and LPC parts of the system are generally assumed time-invariant. This is in contrast to waveform coders such as adaptive transform coding which have greater flexibilities in allocating greater quantization resources to properties of the signal that are deemed most significant at any given time. Such time-varying strategies seem to be a better match to the flexibilities that the human speaker seems to have in producing natural speech. The purpose of this paper is to study the benefits of time-varying bit allocation in linear predictive coders. There are many ways of realizing variable rate coding for the parameters in these coders. An extreme example of zero bit allocation results when a particular parameter set is not transmitted at all, but replaced by a previous-frame version of it.