Example: lossless waveform coding

• Lossless compression of a speech database using ADPCM-like techniques with a forward predictor (shorten)
• Speech sampled at 16 kHz and 16 bits per sample (TIMIT database)

The variation of compression ratio with predictor order is shown in figure 44.

  
Figure 44: Variation of compression ratio with predictor order for a polynomial and general linear predictor

Note that there is considerable compression without a predictor. This is because the prediction error is non-uniform and therefore may be entropy coded. The distribution of prediction errors is shown in figure 45.

  
Figure 45: The probability density function for the prediction error and the Gaussian and Laplacian distributions