from what i've been taught, the function of dither is to decorrelate noise from the signal. so it seems what you're suggesting is using signal statistics to generate an optimum dither for signal enhancement?
On Sat, Nov 20, 2021 at 6:21 PM Sampo Syreeni <de...@iki.fi> wrote: > On 2021-11-20, Nigel Redmon wrote: > > > It’s an interesting topic, for sure, and I don’t mean to slow it down. > > I know myself too: this is kinda nerdish, and not much of application. > Still, bear with me. ;) > > > I just want to interject a practical comment about it. At this point > > in time, when quite affordable streaming services are sending 24-bit > > uncompressed audio, the advantage of gaining back fidelity from a > > truncated (dithered/reduced word size) channel is rapidly running out > > of life time. > > There is a bit of a problem even with wide channels: noise accumulation. > That which in the analogue world called for a precisely controlled gain > structure, over the whole capture and reproduction chain. > > In the world of digital dither, it surfaces in another way: every time > you process your audio, in order to do it well you have to dither down > to some resolution. When you do it additively, you will typically add a > TPDF dither every step along the line. Using 24-bit signals, it'd take > -- off the top of my head -- some 30 signals added in parallel, for the > dither noise to reach the circa 20-bit audibility threshold we care > about in audio work. > > Of course, if those signals were added in series, as is often the case > in mixing, fewer signals would lead to further problems. > > If on the other hand you went with subtractive dither, end-to-end, it'd > almost be the case that your processing topology in the mixing room, or > the number of sources, would not matter at all. Perfectly subtractively > dithered digital channels after all add to each other much like analogue > AWGN ones do. They don't contribute inherent noise or nonlinear > artifacts by coding *at* *all*. They just AWGN-hiss at the lowest > possible level, equal to the 1/2 bit naïve analysis of a PCM channel. > > (Though I've never seen a thorough analysis of this stuff at the > microlevel. Might be I'm off by tenths of decibels or such.) > > > Even for the highly financed MQA, the compression aspect of needing > > only 16 bits is a questionable selling point, and they have to rely on > > more questionable claims of sound improvement because no one cares > > they can fit it on a CD. > > This is the first time I even bumped into MQA. It seems questionable to > someone like me. Even if I'm a great fan of Craven, especially Gerzon, > and perhaps even Stuart (Meridian's work is second to none). > > In fact in the past I've perhaps talked here or on Sursound about some > of the theory seemingly underlying MQA. About that > time-frequency-phase-tradeoff they seem to be pulling -- it was well > shown in the "critical video" of MQA, where they showed the codec to > lead to something like a minimum phase response, instead of a linear > phase one. > > I believe there is a non-MQA oriented earlier paper by Peter Craven > where this is discussed more thoroughly. And I believe *half* of it. > Because: > > 1) There is no way you can get the measured steady state noise floor of > auditory neurons, at the same time that you get their measured S/N and > dynamics, to reconsile with binaural measurements. There is something > weird going on here, and it cannot be fully linear. > > 2) Even worse, from the early dichotic hearing experiments in the > seventies or so, it seems that ultrasound well into the hundreds of > kilohertz range aids measurably in localisation. You can't *hear* any of > the Fourier components of any continuous tone up there, but if you > filter out such components from an analogue-originated two-hearphone > click-sound, 1-2MHz wide in the baseband, you in fact see steadily > deteriorating lateral location accuracy. Thus, ultrasounds might > contribute to time-domain psychoacoustics, without being heard, and > only being sensed as a part of a "steeper impulse with less rise-time, > in the time-domain". > > There is something else about this which I've talked about in the far > past both in here and on Sursound, but I can't right now bring to mind > what it was. Craven has done this stuff a lot, but I think there's > something else also. > > > Carry on, I like the mental stimulation and this is fun stuff. But > > when we get high definition video on our phones and stream 24-bit > > audio that’s already beyond the limits of electronics to reproduce and > > our ears to detect, I’m not sure what you aim to improve. > > One of the things is metrology/telemetry, which is why -- drunk as fuck > -- I also linked my original post to NASA, ESA, whatnot. ;) > > Not all acoustical signals *can* be relayed at 24 bits. Not from > something like a Pioneer or Voyager probe. In space applications at > least, you'd want your signal to not suffer anything like additive > dither, but to come out at a flat, expectable noise floor, utilizing > every last bit there is, over a very narrowband channel. Which is what > subtractive dither then does. > > In terrestrial audio forensics, every bit counts, too. If you want to > correlate mains hum over a longer time, in order to authenticate in time > a low level audio record, you often need to go well below the noise > floor. So that if if your dithering scheme adds noise, evidence might > not be uncovered because of that 6-12dB or so drop we're talking about. > Not to mention how much easier all kinds of analysis is, when you *know* > your channel is statistically linear to the hilt -- quite a number of > analysis modes fail if we can't assume linearity. > > > Like MQA, something like this would have been a lot more compelling > > 10-20 years ago. > > Not a fan of MQA, and that's not the goal at all. I'm not trying to > somehow enhance a signal, using some sort of inband communication > scheme to drive a codec. I'm trying to subtractively dither arbitrary > signals, so that we don't have to add noise to the utility signal. > > Coding theoretically, I'm then trying to spread out the synchronization > signal that I need, as far as possible, and to derive it from the > entropy in the signal itself. So as not to disturb the statistics of the > utility signal any more than absolutely necessary. > -- > Sampo Syreeni, aka decoy - de...@iki.fi, http://decoy.iki.fi/front > +358-40-3751464 <http://decoy.iki.fi/front+358-40-3751464>, 025E D175 > ABE5 027C 9494 EEB0 E090 8BA9 0509 85C2
