3.4 COMBINATION WAVEFORM/MODEL-BASED CODING

The principal limitation of purely model-based speech coding is that no simple mathematical model can fully capture the dynamic, fine-structure, and expressive subtleties of natural speech. Once the structural limits of a model are reached, allocating additional bits to its parameters produces little further perceptual improvement. In contrast, waveform coders such as pulse-code modulation (PCM) and ADPCM aim to reproduce the speech waveform directly and accurately, but require relatively high bit rates—typically above about 16 kbps—to achieve acceptable subjective quality.

This creates an intermediate bit-rate region, approximately 8 to 16 kbps, in which neither approach alone performs well. Model-based coders in this range suffer from audible artifacts due to model mismatch, while waveform coders are inefficient because they expend bits representing perceptually unimportant signal detail. Hybrid waveform/model-based coders were developed to bridge this gap by combining the strengths of both approaches: the efficiency of parametric modeling with the perceptual fidelity of waveform coding.

In a hybrid coder, the speech signal is first represented by a predictive model—typically based on linear prediction—to capture the slowly varying spectral envelope. The remaining prediction error, or residual, is then encoded using a structured excitation or waveform representation rather than being discarded entirely. Early hybrid systems included adaptive predictive coding (APC) and multipulse excitation (MPE), which transmitted a limited number of excitation pulses together with the parameters of a linear predictive model.

Subsequent refinements, such as regular pulse excitation (RPE) and CELP, introduced analysis-by-synthesis techniques in which candidate excitation signals are evaluated by synthesizing speech and comparing it perceptually with the original signal. By explicitly optimizing the excitation to minimize a perceptually weighted error measure, these coders achieve substantially improved naturalness and robustness at moderate bit rates.

Hybrid waveform/model-based coders now form the basis of most modern narrowband and wideband speech-coding standards. Systems such as ITU-T G.728, G.729, and Adaptive Multi-Rate Wideband (AMR-WB) provide near–toll-quality speech at bit rates between approximately 8 kbps and 16 kbps, while sub-band-based hybrids such as G.722 extend these principles to higher audio bandwidths. Together, these techniques represent the practical convergence of waveform and model-based coding approaches in contemporary speech communication systems.