birdlandbill
Head-Fier
This is further under the hood than I've ever been. Got me looking again into DSP texts, noise shaping, various probability distribution function dithers and other goodies. No pretense at fully understanding, just taking a crack at some of this with MatLab to see darkly through DSP looking glass. Loving it. BTW, just a listener with no connection to any audio device manufacturers save with credit card.A. My guess is that optical at 24b would be best sounding. Try both, and see for yourself; use a test track that has good depth (I use an organ track), and the one that has the best perception of depth plus the warmest tone, is the more accurate/transparent one.
B. No - it's not dithered down to 24 bits but very aggressively 11th order noise shaped. Now noise shaping at 768k with my truncation noise shaper (the truncation noise shaper actually has a different architecture to pulse array noise shapers), allows for -301 dB accuracy (so a -301 dB signal is accurate to +/-0.001dB in amplitude and +/- 0.001 deg in phase) and better than 350dB THD and noise within the audio bandwidth. This is measurably perfect small signal performance, and it's something that, based on my listening tests, is essential for transparent sound quality (that is maintaining depth perception).
When the M scaler was first put together with the Blu 2, I spent a lot of time in listening tests to get the maximum SQ performance (and here the problem is trying to add a digital module without it reducing my perception of sound stage depth), and this is how I came up with the 11th order truncation noise shaping. Now this is something one can do at 768k; but at 44.1 or 48k you cannot do this, so when converting to n bits down to 24 bits than you have to use dither. As part of my ADC project, I wanted dither to sound as good as the noise shaping - so started with rectangular and triangular dither (TPDF). Rectangular sounded the worst, but TPDF was still much inferior to the 768k noise shaping; I then used pseudo Gaussian dither, and this got me much closer to the noise shaper performance compared to TPDF - but it was still no where near as good as the noise shaped performance. Incidentally, if you have a truncation setting for a filter, never switch dither to off, the small signal distortion then becomes huge. If you have the option use Gaussian dither - if this isn't available (it's very rare) then triangular or TPDF.
As far as the Davina project goes, I still have more work to do, in order to improve 44.1k truncation to 24 bits. Perhaps making the pseudo Gaussian more Gaussian may help close the SQ gap.
C. Any change in the digital data requires re-quantization if there is a truncated data residual (the bits that can't be transmitted). And my answer B shows this can only be done at 705/768 when using advanced noise shaping. So even using a -0.001dB change would create problems in depth perception; this is why my standard advice is to keep it source bit perfect - and that ignores the fact that any up-sampling filter is not the same as a WTA filter.
Listening to Joel Ross KingMaker as I write this. Stunning. Best recording of vibraphones (and an excellent ensemble, particularly the drummer) I've ever heard due to musicianship/recording quality and wonderful Qobuz (96/24 signal)->HMS->Wave Storm->Dave->Utopia signal transducing system.
Thank you for sharing and Happy Holidays!
Last edited: