For those of you waiting with bated breath (LOL), here is part II, finally…
The author presents an argument that if instead of normalization after truncation (see part I of this blog for an example of truncation), the digital samples are processed by a digitally implemented inverse-RIAA filter, and then carefully matching the gain, the residual error is greatly reduced. In fact, he claims that the magnitude of error is only on the order a few LSBs out of 24 bits, and he claims that this proves to be over four bits less than expected 6 or 7 bit loss.
I have no way of knowing what filtering or algorithm is applied in DSP so I can’t comment on his findings, however today’s floating point DSPs have incredible amounts of processing power, and the accuracy of the computations involved in not in question.
The other argument presented was based on the review of perhaps thousands of available vinyl recording, his own and others. He selected around 40 recordings that were singled out based on the amount of high frequency content. His theory is that if a recording has a high level of high frequency or treble content, the resolution of the bass frequencies would be the most impaired.
This stems from the fact that when doing a digital recording (or an analog to digital conversion), you will eventually run out of headroom, and any transient that exceeds 0dBFS (full-scale) will be clipped.
When doing an inverse-RIAA filter digitally it is imperative that the maximum level on the recording is known, otherwise you will either 1) clip the signal on the peaks if the level is too high or, 2) underutilize the available resolution if the level is too low.
By sampling these specially chosen recordings, and using peak responding, peak-hold audio analyzer software, he discovered that the peaks were lower that what would be expected based on the RIAA curve alone (which would imply that you would have several of resolution for free).
That is not strictly the case however, since the analog gain level must be set based on the peak value of the coherently summed signal of bass, mids, and treble, not just the peak value at discreet frequencies across the spectrum. But remembering that the mids and especially the bass are de-emphasized, and by analysis of the program peak level on the selected recodings, validated the use of the peak responding, peak-hold analysis, and that with LP recordings this is unlikely to exceed 1 bit (6dB).
His argument continues that RIAA equalization in the analog domain followed by digital conversion would be a reduction in fidelity since 1) the program peak level would now be defined by the recording’s bass level, 2) there would now be less resolution available for the treble, so the treble will be truncated instead of the bass, 3) humans have better hearing in the treble range, 4) treble doesn’t benefit from the reconstruction effects of the filter.
So there you go… If there are errors in getting his point across they are probably mine.