|
|
Post by chessparov on Jan 9, 2024 23:01:39 GMT -6
Oh...just to see if I can say this here without toxic responses ensuing... First, as many will be happy for me to say, you might as well dither any linear bit reduction (by "linear", I'm not yet addressing encoding to lossy formats such as mp3). Silly not to, you're exporting anyway, and you'll stave off the potential of an obnoxious form of distortion. A little more info that will set some people off: When exporting, we really only reduce to 16-bit, or from floating point to 24-bit these days. At 16-bit, the threat is real, at 24-bit it's not, but just dither that export in either case and don't fret. As for possible downside: For 16-bit, the added noise floor of dither is very slight, next to impossible to notice without "cheating" to hear the difference (cranking a quiet passage to compare with/without), but it removes the possibility of the far-worse truncation distortion. Dither it. For 24-bit, it doesn't matter either way—impossible to hear either the added noise from the dither, impossible to hear the truncation distortion. The noise floor of any DAC or power amplifier ever made, or yet to be made, dwarfs either. So, if I say "do it anyway", why make people mad and tell the truth that it doesn't matter at 24-bit? Because sometimes it's not so easy to dither, and trying to "do the right thing" (as perceived) has a much larger chance of making a mistake and making the signal worse. So, I believe people should understand, and make the choice for themselves, and be get unnecessarily frightened into doing something that can't possibly matter. Mainly, that case comes up when people who are recording and mixing audio material in a computer like to "send" channels out to external analog processing gear—vintage compressors, for instance, on drums and vocals. Usually it's just a few, but in one case of someone asking, they needed to send to every channel of a large-format mixer, which would require 48 instances of dither plugins running on their computer. As for mp3 and the like, feed the converter what it needs. I think in most cases that's 32-bit flat these days—float never required dither. (Yes, people will even argue this, I'm just answering the question. Choose who to believe, or ask me more about why, if you want to know.) Of course, it you have a "vintage" converter that wants to be fed 16-bit file, dither to 16-bit first. But...just get a modern converter.  I'll just give the most basic reason dither at 24-bit doesn't matter, though there are more reasons. And people who aren't physicists or electrical engineers have a really rough time with this one: It's not possible to make a 24-bit converter at audio levels. We don't make 24-bit converters because we need them or because we are capable of making them. We make them because 24 is one byte more than 16. Basically, pro (+4 dBu balanced) max level is 4 Vrms, consumer (-10 dBV) is 2 Vrms. Let's start with the best case—pro. 24 bits yields 16777216, the smallest step being the reciprocal of that. So, divide 4 Vrms by 16777216 to get the smallest voltage step at full volume: 0.0000002384 Vrms, about a quarter of a microvolt—a fraction of a millionth of a volt. Another way that probably more convenient for audio engineers to look at it is that the bottom bit is about -144 dB from full scale output of the DAC (20 * log10(Power(2,-24)) = -144.49). But if you read specs on DAC (and preamps and amplifiers for the matter), the very best have noise figures at about -120 dB, with a select few inching closer towards -130 but falling well short. That's because even ONE lowly resistor at values required and at audio bandwidths emit noise at around -130, -131 dB. This is thermal noise, called Johnson-Nyquist noise. There are other noise effects down near the bottom of 24-bit, but this is the loudest one and is unavoidable, so we can end the discussion there. And cryogenics is not the answer, because even if the device didn't fail, at some point you'll have to run it into an amp and to speakers and hear it without being frozen to death, in which case it doesn't matter whether the audio is dithered anyway.  BTW, amps are typically far noisier than DACs, because power isn't easy. I know this is incredibly annoying to some. But I believe if you give people the truth, they can make informed decisions. If that decision is to instantiate 48 dither plugins to "be sure" and feel comfortable, yes, do what makes you feel comfortable. But I would not want to be the guy for whom that just won't work, so I don't do it, but fear all the time that I've damaged the enjoyability of my music for some people because I skipped that step. No, it's not possible to hear the difference—now just do what you want to do. Also, I've seen someone post, "for years I haven't dithered my 24-bit file because I didn't know about dither—have I screwed my whole library of music??" Personally, I want them to know that they don't need to lose sleep. Could anyone please give me the DETAILED explanation? I've searched both "hither" and "dither". BTW I seem to recall "Toxic Responses" opening for The Germs and Anthrax in the 90's. The energy in that Mosh Pit, was simply infectious. Unfortunately there was no Cure to follow. |
|
|
|
Post by gravesnumber9 on Jan 10, 2024 0:02:47 GMT -6
Maybe I’m not familiar with the Jets-vs-sharks nature of dither discussion…but you’re welcome to post more any time. Don't you know? When you dither you dither all the way from your first guitar fret till your last dying wav. C'mon. Everybody knows that. |
|
|
|
Post by earlevel on Jan 10, 2024 3:23:02 GMT -6
From Bob and Paul Frindle collectively, and probably others, I’ve been told to always Triangle dither when leaving the box for outboard processing, or, exporting to a lower bitrate, including 24-bit files that have been mixed in 32-bit float and will be going out to 16 *or* 24 for export. Yes, this is sound advice (no pun intended, but hey, it's there). Flavor of dither is personal. But I agree with "stick with triangular" (TPDF). Rectangular isn't decorrelated with with power, but the more apparent problem is that if you pick the ideal (quieter) amount of it, then in total silence you'll have no noise, but the noise will start with the slightest level. In other words, you have noise modulation, and the only cure is to use more dither (and hence noise floor) than you would otherwise need. Triangular and rectangular sound the sound to the ear (white), but triangular is clearly superior for dither, and there 's no way to screw it up as long as you have (at least) ±1 bit of it. (iZotope's default is slight less, but that typically works out. But since ±1 bit is not loud at 16-bit and 24-bit, I don't quite see the point of it, but neither do I think it's a terrible idea. IIRC correctly, I think the iZotope setting for ±1 lsb TPDF is called "Strong".) The shaped dithers are seductive, because they more the noise away from where the ear is sensitive to it, effectively giving you a much deeper apparent bit depth. However, there are some cases where this works against you. At the same time, TPDF may be hiss, but it's pleasant hiss. And even at 16-bit, you're amp may well be worse. And even if it isn't, you'll only hear the hiss easily at extreme monitoring levels between songs. I'm happy for people to disagree on this point. And it's a badge of honor for manufacturers to tout their own proprietary shaped dither. And it's seductive to say that you can squeeze something like 120 dB dynamic range out of 16-bit audio with it. But we don't have that kind of practical* hearing range, and more importantly we don't need it for a good listening experience. (*People will say, "but our hearing has 140 dB dynamic range!" Sure, but not if you intend to hear again. But it's OK, your playback system doesn't have that range either.) Also, people get mislead by the claim "it lets you hear 20 dB below the noise floor". First, if you can't hear a signal because it's too quiet, you still won't hear it. Also, the noise floor got even higher somewhere else, it's just that your ear isn't so good at hearing in that range. Noise-shaped shaping does amazing things for 8-12 bit audio. But the minimum audio depth is 16-bit for a long time now, and at 16-bit, TPDF does the job with room to spare, while having no possibility of adverse affects (assuming the algorithm isn't broken). So there isn't an overwhelming need for this type of noise shaping. And the funny thing is that you're already effectively flawless with a little noise shaping, yet manufacturers tout their particular flavor as bing 1 dB better than the others, when all of them are already beyond people's ability to tell...except that the more extreme you get, the more chance their is for negative side effects. I find this type of noise floor race to be a little humorous. I know people who noise shaped dither, and I'm not trying to tell them it's wrong—and I'm certainly not saying you can't hear the difference. If you know enough to prefer it, use it. My message is for the person asking "how do I do this right, and be certain I'm not mucking thing up?" TPDF is the answer. You'll be right to use it from day one, you'll be right to use in for the next 20 years (16-bits is probably ancient history before then), and if you learn to love shaped dither, you'll be right to use that if you prefer it. |
|
Deleted
Deleted Member
Posts: 0
|
Post by Deleted on Jan 10, 2024 8:39:53 GMT -6
From Bob and Paul Frindle collectively, and probably others, I’ve been told to always Triangle dither when leaving the box for outboard processing, or, exporting to a lower bitrate, including 24-bit files that have been mixed in 32-bit float and will be going out to 16 *or* 24 for export. Yes, this is sound advice (no pun intended, but hey, it's there). Flavor of dither is personal. But I agree with "stick with triangular" (TPDF). Rectangular isn't decorrelated with with power, but the more apparent problem is that if you pick the ideal (quieter) amount of it, then in total silence you'll have no noise, but the noise will start with the slightest level. In other words, you have noise modulation, and the only cure is to use more dither (and hence noise floor) than you would otherwise need. Triangular and rectangular sound the sound to the ear (white), but triangular is clearly superior for dither, and there 's no way to screw it up as long as you have (at least) ±1 bit of it. (iZotope's default is slight less, but that typically works out. But since ±1 bit is not loud at 16-bit and 24-bit, I don't quite see the point of it, but neither do I think it's a terrible idea. IIRC correctly, I think the iZotope setting for ±1 lsb TPDF is called "Strong".) The shaped dithers are seductive, because they more the noise away from where the ear is sensitive to it, effectively giving you a much deeper apparent bit depth. However, there are some cases where this works against you. At the same time, TPDF may be hiss, but it's pleasant hiss. And even at 16-bit, you're amp may well be worse. And even if it isn't, you'll only hear the hiss easily at extreme monitoring levels between songs. I'm happy for people to disagree on this point. And it's a badge of honor for manufacturers to tout their own proprietary shaped dither. And it's seductive to say that you can squeeze something like 120 dB dynamic range out of 16-bit audio with it. But we don't have that kind of practical* hearing range, and more importantly we don't need it for a good listening experience. (*People will say, "but our hearing has 140 dB dynamic range!" Sure, but not if you intend to hear again. But it's OK, your playback system doesn't have that range either.) Also, people get mislead by the claim "it lets you hear 20 dB below the noise floor". First, if you can't hear a signal because it's too quiet, you still won't hear it. Also, the noise floor got even higher somewhere else, it's just that your ear isn't so good at hearing in that range. Noise-shaped shaping does amazing things for 8-12 bit audio. But the minimum audio depth is 16-bit for a long time now, and at 16-bit, TPDF does the job with room to spare, while having no possibility of adverse affects (assuming the algorithm isn't broken). So there isn't an overwhelming need for this type of noise shaping. And the funny thing is that you're already effectively flawless with a little noise shaping, yet manufacturers tout their particular flavor as bing 1 dB better than the others, when all of them are already beyond people's ability to tell...except that the more extreme you get, the more chance their is for negative side effects. I find this type of noise floor race to be a little humorous. I know people who noise shaped dither, and I'm not trying to tell them it's wrong—and I'm certainly not saying you can't hear the difference. If you know enough to prefer it, use it. My message is for the person asking "how do I do this right, and be certain I'm not mucking thing up?" TPDF is the answer. You'll be right to use it from day one, you'll be right to use in for the next 20 years (16-bits is probably ancient history before then), and if you learn to love shaped dither, you'll be right to use that if you prefer it. This but those auto-blank functions in dithers should be turned off at 24-bit and 16-bit triangular where there's not any noise shaping with added high frequency energy. I just leave a 24-bit dither on all the time on my master outs and mute the speakers. Put all of the fancy pro adcs and dacs back to line level, take unweighted measurements, and they have SNR close to really good, modern ic based analog gear if they're any good. Unfortunately the average multichannel converter is still severely compromised and the good ones (current DAD or DAD rebranded as Avid, Lynx, Apogee, Burl, the Dangerous Convert 8, the new Prism Dream etc) are very expensive. Other manufacturers are cheaper and when you put them at line level, usually their inputs or outputs do not spec out well at all from cheap circuitry when you put tones beyond 1 kHz into them. |
|
|
|
Post by jeremygillespie on Jan 10, 2024 9:37:00 GMT -6
I’m just happy I can take this new knowledge and blame my old shit mixes on not dithering. Thanks guys!
|
|
Deleted
Deleted Member
Posts: 0
|
Post by Deleted on Jan 10, 2024 10:40:33 GMT -6
Oh...just to see if I can say this here without toxic responses ensuing... First, as many will be happy for me to say, you might as well dither any linear bit reduction (by "linear", I'm not yet addressing encoding to lossy formats such as mp3). Silly not to, you're exporting anyway, and you'll stave off the potential of an obnoxious form of distortion. A little more info that will set some people off: When exporting, we really only reduce to 16-bit, or from floating point to 24-bit these days. At 16-bit, the threat is real, at 24-bit it's not, but just dither that export in either case and don't fret. As for possible downside: For 16-bit, the added noise floor of dither is very slight, next to impossible to notice without "cheating" to hear the difference (cranking a quiet passage to compare with/without), but it removes the possibility of the far-worse truncation distortion. Dither it. For 24-bit, it doesn't matter either way—impossible to hear either the added noise from the dither, impossible to hear the truncation distortion. The noise floor of any DAC or power amplifier ever made, or yet to be made, dwarfs either. So, if I say "do it anyway", why make people mad and tell the truth that it doesn't matter at 24-bit? Because sometimes it's not so easy to dither, and trying to "do the right thing" (as perceived) has a much larger chance of making a mistake and making the signal worse. So, I believe people should understand, and make the choice for themselves, and be get unnecessarily frightened into doing something that can't possibly matter. Mainly, that case comes up when people who are recording and mixing audio material in a computer like to "send" channels out to external analog processing gear—vintage compressors, for instance, on drums and vocals. Usually it's just a few, but in one case of someone asking, they needed to send to every channel of a large-format mixer, which would require 48 instances of dither plugins running on their computer. As for mp3 and the like, feed the converter what it needs. I think in most cases that's 32-bit flat these days—float never required dither. (Yes, people will even argue this, I'm just answering the question. Choose who to believe, or ask me more about why, if you want to know.) Of course, it you have a "vintage" converter that wants to be fed 16-bit file, dither to 16-bit first. But...just get a modern converter. I'll just give the most basic reason dither at 24-bit doesn't matter, though there are more reasons. And people who aren't physicists or electrical engineers have a really rough time with this one: It's not possible to make a 24-bit converter at audio levels. We don't make 24-bit converters because we need them or because we are capable of making them. We make them because 24 is one byte more than 16. Basically, pro (+4 dBu balanced) max level is 4 Vrms, consumer (-10 dBV) is 2 Vrms. Let's start with the best case—pro. 24 bits yields 16777216, the smallest step being the reciprocal of that. So, divide 4 Vrms by 16777216 to get the smallest voltage step at full volume: 0.0000002384 Vrms, about a quarter of a microvolt—a fraction of a millionth of a volt. Another way that probably more convenient for audio engineers to look at it is that the bottom bit is about -144 dB from full scale output of the DAC (20 * log10(Power(2,-24)) = -144.49). But if you read specs on DAC (and preamps and amplifiers for the matter), the very best have noise figures at about -120 dB, with a select few inching closer towards -130 but falling well short. That's because even ONE lowly resistor at values required and at audio bandwidths emit noise at around -130, -131 dB. This is thermal noise, called Johnson-Nyquist noise. There are other noise effects down near the bottom of 24-bit, but this is the loudest one and is unavoidable, so we can end the discussion there. And cryogenics is not the answer, because even if the device didn't fail, at some point you'll have to run it into an amp and to speakers and hear it without being frozen to death, in which case it doesn't matter whether the audio is dithered anyway. BTW, amps are typically far noisier than DACs, because power isn't easy. I know this is incredibly annoying to some. But I believe if you give people the truth, they can make informed decisions. If that decision is to instantiate 48 dither plugins to "be sure" and feel comfortable, yes, do what makes you feel comfortable. But I would not want to be the guy for whom that just won't work, so I don't do it, but fear all the time that I've damaged the enjoyability of my music for some people because I skipped that step. No, it's not possible to hear the difference—now just do what you want to do. Also, I've seen someone post, "for years I haven't dithered my 24-bit file because I didn't know about dither—have I screwed my whole library of music??" Personally, I want them to know that they don't need to lose sleep. It does matter at 24-bit. The Johnson noise is not random. Reverb cue and notes can disappear. All sorts of things that are heavily distorted and mostly noise like synths, guitar amps, old reverbs will have major playback issues at undithered 24-bit that will impact your mix moves. Notes will disappear into the background and you might ride them up to be audible but the easier solution is to just not distort them in the first place. The quantization error is correlated to the signal in 32-bit float to 24-bit fixed without a final 24-bit dither always for audio or non normalized floating point numbers. Now it's time for extreme nerd shit jmooseEvery 24-bit signal can be stored in a 32-bit float container but the converse is not true. Almost all 32-bit float number quantize to 24-bit fixed numbers. Only normalized 32-bit floating point numbers will come out intact without truncation. That means the integer part of the decimal part, not the mantissa has to be a single digital whole number, usually 1. Floating point numbers in a computer do not have to be normalized. An audio signal will never be all normal floating point numbers. So dither is necessary when converting to fixed point even if a 24-bit sample will pass through a 32-bit float conversion intact if nothing is done (like in most Exclusive mode Core Audio driver interfaces despite what the internet tells you)
Also keep in mind that floating point numbers do not need to be dithered because they cannot be and they round instead of truncate. they round half to even in a computer to prevent bias, which in audio becomes DC. en.wikipedia.org/wiki/Rounding#Rounding_half_to_evenI've never heard rounding distortion if there is enough precision for the function, e.g. I've heard direct form biquads blow up in 32-bit float and zip around when modulated in 64-bit float but good luck hearing better filter structures (or filter structures used with enough precision they won't blow up in real world use or cause noticeable low frequency artifacts) in 64-bit float versus rounded to 32-bit float if for example in a daw that supports them, you choose VST3 which can natively support 64-bit float i/o versus AU, which is limited to 32-bit float i/o. Fixed point math is crazy hard if not impossible and the processors that were clean/useful that used it (excluding old reverbs and delays that were holistically written to sound utilitarian and cool) were generally written by programming gods. Some of those algorithms are still the among the handful of good processes in their respective categories today.
|
|
|
|
Post by jeremygillespie on Jan 10, 2024 11:06:35 GMT -6
I don’t think I’ve ever opened a session at 32float. What am I missing?
|
|
|
|
Post by notneeson on Jan 10, 2024 11:11:27 GMT -6
I don’t think I’ve ever opened a session at 32float. What am I missing? I think it’s a phenomenon purely driven by FOMO. |
|
|
|
Post by gravesnumber9 on Jan 10, 2024 11:24:39 GMT -6
I don’t think I’ve ever opened a session at 32float. What am I missing? I think it’s a phenomenon purely driven by FOMO. Studio One at least (and I think Reaper) default to mix engine at 32 bit float. Logic does now as well I believe and probably others I'm sure. So not recording at 32 bit float, but under the hood processing. Different thing than recording at 32 bit. |
|
|
|
Post by jmoose on Jan 10, 2024 12:02:47 GMT -6
I don’t think I’ve ever opened a session at 32float. What am I missing? Nothing at all. For one thing, as we established earlier there are no true 32 bit converters on the market. So... well, there 'ya go right? I absolutely think some people are confusing the internal processing/word length of the DAW... aka 32/64 bit mix engines with the actual audio which, at best case scenario is 24 bit. Not the same thing. And what's kinda scary is that almost everything being talked about here is like, audio student 101 digital theory... aka "how things work" |
|
|
|
Post by notneeson on Jan 10, 2024 12:06:31 GMT -6
I think it’s a phenomenon purely driven by FOMO. Studio One at least (and I think Reaper) default to mix engine at 32 bit float. Logic does now as well I believe and probably others I'm sure. So not recording at 32 bit float, but under the hood processing. Different thing than recording at 32 bit. Yeah, I’m talking more about folks who output 32bit masters for future proofing. Good clarification! |
|
|
|
Post by geoff738 on Jan 10, 2024 12:19:41 GMT -6
I was told there would be no math.
Cheers,
Geoff
|
|
Deleted
Deleted Member
Posts: 0
|
Post by Deleted on Jan 10, 2024 13:13:19 GMT -6
I don’t think I’ve ever opened a session at 32float. What am I missing? Every modern DAW is processing in 32-bit and 64-bit float.
All modern converter ics are hopefully running at 2 to 6 bits at ultra high sample rates with tons of noise shaping and digital filtering. Yes they are all automatically dithering.
The AD converter chips usually output 24 bit fixed point pcm but maybe have 20ish bits of real world precision.
Whether you choose to have your daw write this as 24-bit fixed or 32-bit float or even 64-bit float is up to you but the moment it modifies that 24-bit fixed audio, it will almost always be in 32-bit or 64-bit float until you either render it as a fixed-point file or it is sent back out to your converters (almost always at 24-bit eventually) but most of them do not automatically dither it because the people writing the daws are not as mathematically wiz bang as the people creating the converters, they don't want to waste cpu cycles, or they expect you to do it yourself based on the output format and the bit depth of your playback converter, which can be anywhere from 14 to 24-bits. Yes there were common 14-bit converter ICs.
The mostly stand alone recorders that claim to be able to output 32-bit float are doing so by having two 24-bit converters at different gains and choosing one to create the 32-bit floating point sample that it saves to disk. This usually causes some crossover distortion somewhere and you can still clip the analog front end or the higher gain converter IC inputs and save it. There is a lot of idiot proofing in the Sound Devices and Zoom recorders for sound guys who cannot gain stage anything using their built in mixers and trims.
The DA converter chips accept 24-bit fixed point. They will truncate if not dithered. The ESS, recent AKM, recent TI ICs have 32-bit processing on them and some can accept 32-bit audio but again they only have about 20ish bits of precision in the real world.
|
|
|
|
Post by Bob Olhsson on Jan 10, 2024 13:15:43 GMT -6
|
|
Deleted
Deleted Member
Posts: 0
|
Post by Deleted on Jan 10, 2024 13:30:10 GMT -6
I don’t think I’ve ever opened a session at 32float. What am I missing? Nothing at all. For one thing, as we established earlier there are no true 32 bit converters on the market. So... well, there 'ya go right? I absolutely think some people are confusing the internal processing/word length of the DAW... aka 32/64 bit mix engines with the actual audio which, at best case scenario is 24 bit. Not the same thing. And what's kinda scary is that almost everything being talked about here is like, audio student 101 digital theory... aka "how things work" It's not even 24-bit at best. Building a true 24-bit 20hz to 20khz converter is physically impossible. Johnson-Nyquist noise rises with frequency, impedance, and heat to say the least. High frequency tones are already higher in lowest theoretically possible noise at room temperature than 24-bit dithered pcm. Let's not even get into noise of mics amplified to line level with nice sounding equipment that runs hot. Most younger audio students, sound guys, and laymen have no idea how things work. Even how sampling works or how the better digital processors (and now converters) were made to have almost equivalent performance at any sample rate even going back to the Waves Renaissance and Sony Oxford days. Even for modern developers, there are still things like unnecessary upsampling, poorly chosen filters, lack of functional detectors and meters, lack of output dither, or as a long term KVR poster found out www.logicprohelp.com/forums/topic/116014-apparently-logic-dithers-after-truncation/ eek |
|
Deleted
Deleted Member
Posts: 0
|
Post by Deleted on Jan 10, 2024 13:31:46 GMT -6
Yeah for freezes and renders typically. No reason to truncate to 24-bit when you can freeze or render at 32-bit float. Unless you've been monitoring through a 24-bit dither and want to print that dither, when there's no point in printing 32-bit float. |
|
|
|
Post by chessparov on Jan 10, 2024 14:10:24 GMT -6
Geez Bob. No "Wonder"  you passed the Motown IQ test portion, as a kid! Chris |
|
|
|
Post by ab101 on Jan 10, 2024 15:56:50 GMT -6
So, quite a bit of this is over my head -
It seems clear that one should dither from 24bit to 16 bit.
Once dithered to 16 bit, should dither be applied again to go to MP3? Or does the dither from 24 to 16 count for that?
Sorry for my ignorance on this.
|
|
|
|
Post by Bob Olhsson on Jan 10, 2024 16:30:55 GMT -6
My understanding is there is no need. I suspect the ideal way might be to use the proper amount of dither for a 16-bit file but then not truncate into the MP3 or AAC file and let the player output handle the truncation. I'm just not sure how you would do it.
|
|
|
|
Post by earlevel on Jan 10, 2024 17:28:56 GMT -6
[It does matter at 24-bit. The Johnson noise is not random. Reverb cue and notes can disappear. All sorts of things that are heavily distorted and mostly noise like synths, guitar amps, old reverbs will have major playback issues at undithered 24-bit that will impact your mix moves. Notes will disappear into the background and you might ride them up to be audible but the easier solution is to just not distort them in the first place. The quantization error is correlated to the signal in 32-bit float to 24-bit fixed without a final 24-bit dither always for audio or non normalized floating point numbers. "Not random"? It's effectively white in the audio band and far beyond, and Gaussian in the bandlimited sense as well. I'm not sure what your point is here, so I'll leave it at that. I wasn't implying it can replace dither, largely because it's possible to use DSP to generate or process signals so that there isn't Johnson noise present (or not enough), nor was I implying strictly masking with Johnson noise in DACs and subsequent electronics (probably true, just not addressing that). What I was mainly getting at is that the bottom few bits of 24-bit audio can't be played back with any accuracy whatsoever, yet we are talking about dithering the very bottom so it's "correct". Even if we could hear down there. On the issues of signals disappearing, with respect to 24-bit dither...OK, I just disagree that this is a problem, but hard to discuss since I have no idea under what scenarios you imagine this would happen, considering that anything low enough to disappear would be insanely low level but then bought back up with an insane amount of gain or you wouldn't hear it...and at that point the only difference between that and dithered is that you'd do those unlikely steps and the result would be buried in "hiss". I'm not sure if you're counting that as a win over disappearing, but why would you take those steps to undermine the audio in the first place? Clearly, I'm not understanding your point here. I want to make it clear that I'm not trying to "win" on the importance of dithering 24-bit. It's more like I make my case, you make your case, and whomever wants to be influence by either is fine with me. But at least they might notice there is not universal agreement on the point. (It doesn't mean I won't defend the validity of individual point I make, though. ) LOL, a bit of exaggeration on fixed point, I think, as someone with considerable experience in fixed point DSP. Sure, direct form biquads have issues, largely due to quantization of the poles (where the effects are usually at low frequencies), and will blow up when those poles are close to the unit circle. But with a different filter structure (or often simply error shaping), that's not an issue. Yes, I've see DSP coded by people who don't understand this, making some product feature marginally unstable. True that with floating point, the implementor can be blissfully ignorant of these kind of issues and usual get by fine. It won't save you from everything, but if people stick to 64-bit float and its huge mantissa, it's very unlikely to get bit by even pitfalls like summing small to large. |
|
|
|
Post by earlevel on Jan 10, 2024 17:42:41 GMT -6
So, quite a bit of this is over my head - It seems clear that one should dither from 24bit to 16 bit. Once dithered to 16 bit, should dither be applied again to go to MP3? Or does the dither from 24 to 16 count for that? Sorry for my ignorance on this. I don't have broad experience with many mp3 encoders, so I have to make hat I think are reasonable assumptions. The main one is that any modern mp3 encoding software will accept floating point. At very minimum, 24-bit, but I think probably 32-bit float. If so, there is no need to dither. If you need to make a 16-bit audio file from 24-bit or float32, dither it. If you also need to make an mp3, make it from the 24/32 source, not the 16-bit. If you somehow have an archaic mp3 encoder that requires 16-bit audio input, use the dithered file. If you truncate to 16-bit without dither, you bake in potentially correlated (count on it) truncation error, and the mp3 encoder will treat it as it would a musical signal. In general, stay with your maximum bit depth as long as you can. If you truncate, dither. Yes, I do say that it doesn't matter with 24-bit, not because I advise against it, but there are special situations where knowing it doesn't matter can save you a lot of grief—like the guy who is concerned that he needs 48 channels of dither plugins in order to mix it with an outboard analog mixer. Finally, if the mp3 converter is a good one, trust that the developer also understands the issues of word length reduction. You already trust your plugins to not wreck your audio, there is no reason to be especially paranoid of mp3 encoders. Especially considering that you already know that they are maiming your source material–the nature of lossy encoding.  |
|
|
|
Post by Bob Olhsson on Jan 10, 2024 17:52:05 GMT -6
The problem is common 16-bit decoders and not the encoder. Dithering is not difficult for anybody, it's just that lots of developers have shown themselves to be lazy and I tend to not want to trust them.
|
|
Deleted
Deleted Member
Posts: 0
|
Post by Deleted on Jan 10, 2024 17:57:27 GMT -6
[It does matter at 24-bit. The Johnson noise is not random. Reverb cue and notes can disappear. All sorts of things that are heavily distorted and mostly noise like synths, guitar amps, old reverbs will have major playback issues at undithered 24-bit that will impact your mix moves. Notes will disappear into the background and you might ride them up to be audible but the easier solution is to just not distort them in the first place. The quantization error is correlated to the signal in 32-bit float to 24-bit fixed without a final 24-bit dither always for audio or non normalized floating point numbers. "Not random"? It's effectively white in the audio band and far beyond, and Gaussian in the bandlimited sense as well. I'm not sure what your point is here, so I'll leave it at that. I wasn't implying it can replace dither, largely because it's possible to use DSP to generate or process signals so that there isn't Johnson noise present (or not enough), nor was I implying strictly masking with Johnson noise in DACs and subsequent electronics (probably true, just not addressing that). What I was mainly getting at is that the bottom few bits of 24-bit audio can't be played back with any accuracy whatsoever, yet we are talking about dithering the very bottom so it's "correct". Even if we could hear down there. On the issues of signals disappearing, with respect to 24-bit dither...OK, I just disagree that this is a problem, but hard to discuss since I have no idea under what scenarios you imagine this would happen, considering that anything low enough to disappear would be insanely low level but then bought back up with an insane amount of gain or you wouldn't hear it...and at that point the only difference between that and dithered is that you'd do those unlikely steps and the result would be buried in "hiss". I'm not sure if you're counting that as a win over disappearing, but why would you take those steps to undermine the audio in the first place? Clearly, I'm not understanding your point here. I want to make it clear that I'm not trying to "win" on the importance of dithering 24-bit. It's more like I make my case, you make your case, and whomever wants to be influence by either is fine with me. But at least they might notice there is not universal agreement on the point. (It doesn't mean I won't defend the validity of individual point I make, though. ) LOL, a bit of exaggeration on fixed point, I think, as someone with considerable experience in fixed point DSP. Sure, direct form biquads have issues, largely due to quantization of the poles (where the effects are usually at low frequencies), and will blow up when those poles are close to the unit circle. But with a different filter structure (or often simply error shaping), that's not an issue. Yes, I've see DSP coded by people who don't understand this, making some product feature marginally unstable. True that with floating point, the implementor can be blissfully ignorant of these kind of issues and usual get by fine. It won't save you from everything, but if people stick to 64-bit float and its huge mantissa, it's very unlikely to get bit by even pitfalls like summing small to large. the distortion is correlated to the signal and affects higher level signals and perception. clean recordings, distant and room mics, and instruments that are mostly noise and distortion like high gain guitars are most affected. most direct form biquads cannot be automated safely even at 64-bit float and with ramped parameter changes. state variable filters fix that and get rid of many artifacts too that result in clearer sound. there's still poor fixed point dsp in many digital products (certain brands of interfaces and consoles for one) but it's absent from all major daws and plugins now outside of some egregious examples where users found fixed point truncation in a certain developer's products whose former owner would ban people them for reporting such bugs in official channels. |
|
|
|
Post by Bob Olhsson on Jan 10, 2024 18:01:23 GMT -6
Geez Bob. No "Wonder" you passed the Motown IQ test portion, as a kid! Chris I was one of the dumber people at Motown but I wouldn't trade the experience and what I learned for anything. I've always preferred being the dumbest person in the room. I only understood maybe 10% of the conversations in JJ's newsgroup but I saw digital as on the verge of replacing analog. |
|
|
|
Post by earlevel on Jan 10, 2024 19:03:30 GMT -6
examples: room mics, sends to outboard reverbs, recordings of distorted synths and guitars. commercial music? warp records, shoegaze, van halen, swedish death metal, alan holdsworth with the synth axe, emperor, older classical records especially from 50-80s, lofi old live jazz recordings, various 80s to early 2000s cds with the same mastering but only 1 region is dithered. there's lots of examples. change volume in floating point, throw on a 24-bit dither, output them to 24-bit drivers or converter, and they have clearer sound than if no dither and volume change or if the player does not pass the samples directly in either 24-bit fixed point wordlength or 32 bit floating point container. conversely emulations of old digital equipment or algorithm ports do not sound right because their iconic algorithms were holisticaly tailored to run on primitve dsp chips and arrays with tons of digital distortion. none of the vintage lexicon plugins sound right. neither do the uad ams rmx 16, the turbo paco version of the alesis algorithms, the emts (uad, psp, ddmf, softube) are all over the place on if they work on a specific source, both brands of quantec plugins sound different and they probably are doing close to the same thing, the catalin bread pedals are way off versus the real yamaha fx units, h3000 factory is wonky vs the h9000 plugs and h9 pedal, the valhalla shimmer bloom isn't close to the classic midiverb ii shoegaze sound etc. in stark contrast to the uad2 and weiss native ports from modern sharcs that can null or be cleaner
I'm not following how you "lose things" at < -138 dBFS...even before you run them through the electronics supplying noise at a substantially higher level...but OK, I understand your concerns. |
|
BTW, amps are typically far noisier than DACs, because power isn't easy.
you passed the Motown IQ test portion, as a kid! 
