Audio Signal Processing and Coding

7.5 CNET CODERS

Research at the Centre National d'Etudes des Telecommunications (CNET) resulted in several transform coders based on the DFT and the MDCT.

7.5.1 CNET DFT Coder

In 1989, Mahieux, Petit, et al. proposed a DFT-based audio coding system that introduced a novel scheme to exploit DFT interblock redundancy. Nearly transparent quality was reported for 15-kHz (FM-grade) audio at 96 kb/s [Mahi89], except for some highly harmonic signals. The encoder applies first-order backward-adaptive predictors (across time) to DFT magnitude and differential phase components, then quantizes separately the prediction residuals. Magnitude and differential phase residuals are quantized using an adaptive nonuniform pdf-optimized quantizer designed for a Laplacian distribution and an adaptive uniform quantizer, respectively. The backward-adaptive quantizers are reinitialized during transients. Bits are allocated during step-size adaptation to shape quantization noise such that a psychoacoustic noise threshold is satisfied for each block. The perceptual model used is similar to Johnston's model that was described earlier in Section 5.6. The use of linear prediction is justified because it exploits magnitude and differential phase time redundancy, which tends to be large during periods when the audio signal is quasi-stationary, especially for signal harmonics. Quasi-stationarity might occur, for example, during a sustained note. A similar technique was eventually embedded in the MPEG-2 AAC algorithm. ...