CELP model uses an extensive codebook of excitation waveforms, and a closed loop search mechanism for identifying the best excitation in the set for every frame of input speech, without rigidly classifying the input as voiced or unvoiced. The above adaptivety leads to natural sounding speech at relatively low bit rates.Further, if the spectral envelope in the speech model is estimated in a backward manner, based on a recent past history of quantized speech, rather than a current part of unquantized speech, the result is the low delay CELP system of Figure 3. The LD-CELP algorithm has been shown to provide high-quality coding of telephone speech at 16 kbit/s, which turns out to be 2 bits per sample with 8 kHz sampling. With 16 kHz sampling of 7 kHz speech, and a target of 32 kbit/s, the bit rate is again 2 bits per sample. Wideband speech is however, harder to code since the data is highly unstructured at high frequencies and the spectral dynamic range is very high.
Figure 3: (a) Forward (b) Backward adaptive system for CELP