jlg@lanl.ARPA (03/28/85)
> Let's follow a signal through the Philips system. At the beginning, the 2 > LSBs are truncated and applied to the digital filter. The filter > interpolates 3 new data points between samples, `recovering' the lost 2 > bits if 16 were used at the output, which they are not. Thus the base S/N > of the Philips system is 84dB. Since the final bandwidth of the system is > 1/4 the effective bandwidth of the filter output(i.e., the maximum > bandwidth of an input signal sampled at the filter clock frequency) we gain > 6dB in S/N. The roundoff feedback further averages the quantization error > to return to a 96dB S/N. This is not what my article on the Philips system says. It claims that full 16-bit data items are fed to the digital filter with three words of zero between each. The digital filter is a 96 stage discrete convolution integral with multipliers of 12-bits each. The filter keeps all intermediate values to the full 28-bits (product of 16-bit data with 12-bit multiplier). After the digital filter, the 28-bit result is truncated to feed through the 14-bit DAC (This 14-bit result is dithered so that the average of four 14-bit values equals the same result as a single 16-bit filtered value would have been). This is the reason that Philips claims that the 14-bit system looses no information. J. Giles