Subband coding with adaptive prediction for 56 kbits/s audio

This paper reports on a continuation of previous work on the digital coding of 7 kHz wide-band audio at 56 kbits/s [1]. Focus in the present paper is on 5-band subband coding (SBC). Specifically, we consider SBC with adaptive PCM coding in each subband (SBC-APCM); DPCM with backward adaptive prediction (ADPCM); and, most importantly, the combination of the two techniques, SBC-ADPCM. The main conclusion of this paper is that SBC-ADPCM, with a fourth-order backward adaptive predictor in each subband, has a much better objective performance than SBC-APCM for both speech and music inputs, with segmental signal-to-noise ratio gains in the order of 6-8 dB. Informal listening tests confirm that these SNR gains are accompanied by gains in subjective quality. While SNR gains were no higher with two eighth-order predictors that were carefully optimized, small perceptual gains were noticeable with certain inputs. In the absence of transmission errors, there was very little difference between the performance of an 8-pole predictor and a 2-pole 6-zero predictor. In the full-band ADPCM coder, however, the 8-pole predictor was significantly better in the zero-error case.