- 第 14 页 -
原貼的來容真的很豐富, 剛看到有人問類似問題, 故此把其中對話剪了出來
Cameron:
Greatthread folks. I myself do believe that how we deal with headroom in DAW's when mixing is among the most important factors in recording quality.
you guys think the consequences of the 32bit Floating Point file that some DAW are capable ofdealing with factors in to this discussion. I am not sure if folks are aware ofthe fact that on a 32 bit FP DAW you can smash a stereo buss so it clips very audibly,then save a 32bit floating point file of that clipping audio, bring it backinto a program that excepts these files like wavelab/Nuendo/Sequoia bring thesefiles down in amplitude so they do not reach 0dBFS and the distortion willdisappear. When I read this for myself in my Nuendo manual I did not believe ituntil I tried it for myself. It has been years since I felt like I was killingmy stereo master and I I don't consider it anymore a challenge in Nuendo thanit is on our Neve 88R or API.
PAUL :
Yes - this is correct. If all the processesand signal paths are running host floating point (and never reach the outsideworld) it will tolerate overs without actually clipping and reducing the gainafterwards will restore an unclipped signal - providing this is done before hitting any outputs (which are of course fixed point). If you are careful FPentirely host systems can provide (almost accidentally) useful headroom andmore freedom..
Places where this can fall down are if you areusing some 3rd party add on processing cards which may interface via 24bit datastreams, in PT RTAS where the next plug may be TDM (which is a fixed point24bit interface) and of course sending in and out of any outboard gear.
Alsoit's best to tread carefully with plugs that need an operating reference of thereal signal level such as compressors, limiters, maximisers and 'characterprocesses' that use non-linear processing to develop harmonic character. Thesecan suffer (some quite badly) from signals greatly above 0dBfs..
DSM is designed to operate around 0dBfsand the controls (such as threshold) allow sensible operation down to -24dBfsand output gain up to +24dBfs.
Itruns internally at DP floating point and will pass signals of any valid FPlevel without clipping.
However the limiter is aiming at the W/Sstandard program max operating level of 0dBfs (as we have no other choice foroutput media), so engaging the limiter section will constrain the output to 'real'program levels.
Thelimiter section will deal sympathetically with signals up to +6dBfs (whilstlimiting the output to 0dBfs of course) and over-driving up to +6dBfs isencouraged, but beyond +6dBfs it will start to hard limit..
------------------------------------------------------------------------------------------------------------------------------------------------------
Astimony:
Certainly no one doubts your assertion in thisregard.
However, I have read the paper here, itdescribes dithering 32-bit float. It has 3 credible authors, 2 from the University of Budapest and one from Stanford.
I'm a more than competent computer programmer,but I would have a great deal of difficulty implementing a 32 bit dither on thebasis of that paper.
Could you share some pseudo-code with us,describing how to dither a 32 bit float signal? Or some real code, perhaps justthe dither portion from the DSM?
Or, and this is a bit of a guess, when you doplugins do you convert to fixed point, do the processing in fixed point, andthen dither/convert to 32 bit as a last step before the values leave theplugin? I am inferring that from the wording of your post.
It would be great to get this information outthere, in a simple fashion that can be understood easily by plugin and DAWdevelopers. Perhaps then the problem with round off error and quantizationerror leading to inharmonic distortion can be resolved.
If you can't share your methods because theyare trade secrets or because they provide a competitive advantage, that's fineand quite understandable -- given the power of the marketplace.
Take care,
Paul:
There is nothing wrong in these papers andthe situation they describe is correct - from a general point of view..
However for our purposes we are after removingthe harmonic distortion in the legal range of the signal - this harps straightback to several previous posts on float I have made in this thread?
Sobearing in mind that luckily we have a legal real world value of signal rangewhich extends from the noise floor to flat out (+/-1.0) - regardless of whetherit is represented as float or fixed - it is perfectly possible to establish alegal range of signal from say -140dBfs to 0dBfs and remove the quantisationdistortion using dither. By legal we mean one that does not contain unwantedquantisation distortion.
Sofor instance in stuff running float like the DSM (actually running doubleprecision 64bit float) - we still actually dither the output to a real intendedrange of signal (i.e. -140dBfs or so up to 0dBfs) - which means that it will bedistortion free whether feeding 32bit float or 24bit fixed down line :-)
Inother words - regardless of the math representation of the signal, we stillneed to establish a real range if one does not already exist (e.g. as infloat)...
Ok -so what happenes in practice then?
1.The signal comes out of the plug-in in float with a known SNR to flat out range(with dither being generated in the float domain) - which will pass to 32bitfloat or 24bit fixed without quantisation distortion.
Whatit ends up being is determined by the DAW and not us - some will pass full64bit float, some will pass 32bit float, some like TDM, PowerCore and otherswill pass only 24bit fixed. All output formats are actually fixed point 24bits.Where they end up 16bits (i.e. CD etc.) we provide the 16bit dither option,just in case the DSM ends up as the last thing in the signal chain before CDfiles - so all end point eventualities are covered.
2.Signal level values below the noise floor (i.e. in float systems) cannot getbelow the minimum value - because of the added dither - so no problem there anymore.
3.Potentially illegal and out of range end point signal levels above max +/-1.0 =0dBfs will clip on a fixed point target format (CD, AES, 24bit files etc.)
Butwhen feeding a floating target (internal DAW etc.) the signal will continue topass without actually clipping - but with quantisation distortion because above +/-1.0 = 0dBfs the dither no longer matches the scaling caused by the changing float exponent..
So in a float system the headroom created by not clipping above max legal signal comes at a cost.
But the saving grace of the entirely circumstantial float headroom is that the quantisation distortion only occurs at levels above flat out which are veryloud anyway - so the chances of hearing the distortion are vanishingly small:-)
--------------------------------------------------------------------------------------------------------------------------------------------------------
JimJohnston told me that if people are monitoring very loud, they won't be able tohear clipping distortion!
Paul :
From experience I would say that's true. Atvery loud levels the ears actually compress too - and the whole can turn into amush that overwhelms our ability to discern much at all, the HF distortioncaused by clipping just joins the other top end mush we hear...
Andin fact creating this 'mush' at lower levels is actually a large part of whatpasses for production these days - because it can sound 'loud and powerful'even when played at lower levels through a boom box, because it has the 'cuesfor loudness' even though it's quiet. I would say the loudness wars are beingdriven largely by this effect. And of course as designers we have made use ofthis kind of thing in real products :-)
Whenit comes to distortion in general though, I am puzzled by it because peopleseem to vary wildly about how much there can be before they perceive it orcomplain about it. Most modern music releases have far more distortion than Ican listen to - but they are selling to others that are not complaining soloudly - so is it just me I often ask?
Butthen again - are people who tolerate it just conditioned to do so, having neverheard much else? DO we become accustomed to it I wonder? Is it even a form ofdeafness that comes from listening to highly compressed music - even if atlower levels? The ear/brain system is highly complex and adaptive - it could becompensating in some way?
Butone thing I can say from research is that it's amazing how much of certainkinds of distortion can become assimilated into the ear/brain system and thenignored if it's part of some recognisable instrument - that is revealed asawful when attached to a sine wave.
Mytheory is that the ear/nervous system is actually highly non-linear and quitevariable, but the brain cleverly corrects this all the time, allowing us tohear sine waves as pure despite the non-linear biological system. Theperception of a pure sine wave is therefore a special case for the ear/brainsystem, which is more easily disturbed than other circumstances..
Ifyou present this system with distortions of the same kind as it's own inherentones, to a large degree it will correct our perceptions of those too - up to apoint..
Atthe point where these distortions are so bad that the ear/brain system isalmost overwhelmed, if the sound is of a previously known and recognisedinstrument (or something else familiar) it helps the brain to deliver arelatively clean perception of it despite the distortion. And in fact it may bethe case that the system can actually learn to do so more effectively over time- if such distortions are present for long periods on program that is highlyfamiliar, like music.
Inresearch I have done I have been amazed to the degree our ear/brain systems areable to adapt - the adaptation has often been so deep that the removal of astimulus can actually make one feel dizzy, disorientated and quite sick!Incredible stuff I would never have imagined before it happened.
----------------------------------------------------------------------------------------------------------------------------------------------------------
now it makes sense to me.
we can only listen to floating point audiodata thru the fixed point window of a D/A converter... and the dither istailored to this D/A....
Paul :
Not just D/A converter - but the real worldin it's entirety!!
Yes- this is the point. Regardless of the math representation within theapplication itself all signal must re-enter the real world in a limited andbounded range.
Allwe are doing is setting what this range is - so that the end product is suitablydithered - so it has no discernible (or even measurable) quantisation artefact- when represented in it's intended output format.
Itisn't clever particularly at all - to me it's pure common sense?
Thefloating representation is simply equivalent to scientific notation - i.e. anumber of significant figures with a multiplier for the range - i.e. 1.234 x10^3 = 1234.???????????. What this tells us is that however large or small theactual value is - it will only ever be represented by 4 digits. These 4 digitscan be ones, tens, hundreds or thousands. What we have is 4 digits of accuracyonly.
In32bit audio we have (just a little less than) 24bits of reliable accuracy - wedecide where in the scale these will be placed.
Andwe have no choice but to align these with real world signal - which is ofcourse fixed point +/-1.0
Sowe decide the lowest level for this signal wrt what is the highest real worldlegal value - and give it the required statistical uncertainty (i.e. dither -noise floor).
Nowinside the processing we can do whatever we like to make use of the floatingsystem (coeffs bigger than +/-1.0 - values smaller than -140dB or so), but whatemerges in the end must conform to the signal requirements of the outputsystem..
Ifyou chose to ignore these requirements by simply 'going for it regardless' infloat applications then there is a penalty - however luckily n most cases it issmall and only happens on very loud signals that effectively mask the errors..
Where it really breaks down is when anapplication is doing something non-linear on purpose - this non-linearity willhave to be referred to some real world signal level - and that is where nastythings can arise if you ignore the actual signal levels involved..
Thisis what happens if you blast extremely high levels through the DSM (apparentlywithout problem) - and then hit the limiter button.
[ 本帖最后由 himhui 于 12-2-18 14:04 编辑 ]
Cameron:
Greatthread folks. I myself do believe that how we deal with headroom in DAW's when mixing is among the most important factors in recording quality.
you guys think the consequences of the 32bit Floating Point file that some DAW are capable ofdealing with factors in to this discussion. I am not sure if folks are aware ofthe fact that on a 32 bit FP DAW you can smash a stereo buss so it clips very audibly,then save a 32bit floating point file of that clipping audio, bring it backinto a program that excepts these files like wavelab/Nuendo/Sequoia bring thesefiles down in amplitude so they do not reach 0dBFS and the distortion willdisappear. When I read this for myself in my Nuendo manual I did not believe ituntil I tried it for myself. It has been years since I felt like I was killingmy stereo master and I I don't consider it anymore a challenge in Nuendo thanit is on our Neve 88R or API.
PAUL :
Yes - this is correct. If all the processesand signal paths are running host floating point (and never reach the outsideworld) it will tolerate overs without actually clipping and reducing the gainafterwards will restore an unclipped signal - providing this is done before hitting any outputs (which are of course fixed point). If you are careful FPentirely host systems can provide (almost accidentally) useful headroom andmore freedom..
Places where this can fall down are if you areusing some 3rd party add on processing cards which may interface via 24bit datastreams, in PT RTAS where the next plug may be TDM (which is a fixed point24bit interface) and of course sending in and out of any outboard gear.
Alsoit's best to tread carefully with plugs that need an operating reference of thereal signal level such as compressors, limiters, maximisers and 'characterprocesses' that use non-linear processing to develop harmonic character. Thesecan suffer (some quite badly) from signals greatly above 0dBfs..
DSM is designed to operate around 0dBfsand the controls (such as threshold) allow sensible operation down to -24dBfsand output gain up to +24dBfs.
Itruns internally at DP floating point and will pass signals of any valid FPlevel without clipping.
However the limiter is aiming at the W/Sstandard program max operating level of 0dBfs (as we have no other choice foroutput media), so engaging the limiter section will constrain the output to 'real'program levels.
Thelimiter section will deal sympathetically with signals up to +6dBfs (whilstlimiting the output to 0dBfs of course) and over-driving up to +6dBfs isencouraged, but beyond +6dBfs it will start to hard limit..
------------------------------------------------------------------------------------------------------------------------------------------------------
Astimony:
Certainly no one doubts your assertion in thisregard.
However, I have read the paper here, itdescribes dithering 32-bit float. It has 3 credible authors, 2 from the University of Budapest and one from Stanford.
I'm a more than competent computer programmer,but I would have a great deal of difficulty implementing a 32 bit dither on thebasis of that paper.
Could you share some pseudo-code with us,describing how to dither a 32 bit float signal? Or some real code, perhaps justthe dither portion from the DSM?
Or, and this is a bit of a guess, when you doplugins do you convert to fixed point, do the processing in fixed point, andthen dither/convert to 32 bit as a last step before the values leave theplugin? I am inferring that from the wording of your post.
It would be great to get this information outthere, in a simple fashion that can be understood easily by plugin and DAWdevelopers. Perhaps then the problem with round off error and quantizationerror leading to inharmonic distortion can be resolved.
If you can't share your methods because theyare trade secrets or because they provide a competitive advantage, that's fineand quite understandable -- given the power of the marketplace.
Take care,
Paul:
There is nothing wrong in these papers andthe situation they describe is correct - from a general point of view..
However for our purposes we are after removingthe harmonic distortion in the legal range of the signal - this harps straightback to several previous posts on float I have made in this thread?
Sobearing in mind that luckily we have a legal real world value of signal rangewhich extends from the noise floor to flat out (+/-1.0) - regardless of whetherit is represented as float or fixed - it is perfectly possible to establish alegal range of signal from say -140dBfs to 0dBfs and remove the quantisationdistortion using dither. By legal we mean one that does not contain unwantedquantisation distortion.
Sofor instance in stuff running float like the DSM (actually running doubleprecision 64bit float) - we still actually dither the output to a real intendedrange of signal (i.e. -140dBfs or so up to 0dBfs) - which means that it will bedistortion free whether feeding 32bit float or 24bit fixed down line :-)
Inother words - regardless of the math representation of the signal, we stillneed to establish a real range if one does not already exist (e.g. as infloat)...
Ok -so what happenes in practice then?
1.The signal comes out of the plug-in in float with a known SNR to flat out range(with dither being generated in the float domain) - which will pass to 32bitfloat or 24bit fixed without quantisation distortion.
Whatit ends up being is determined by the DAW and not us - some will pass full64bit float, some will pass 32bit float, some like TDM, PowerCore and otherswill pass only 24bit fixed. All output formats are actually fixed point 24bits.Where they end up 16bits (i.e. CD etc.) we provide the 16bit dither option,just in case the DSM ends up as the last thing in the signal chain before CDfiles - so all end point eventualities are covered.
2.Signal level values below the noise floor (i.e. in float systems) cannot getbelow the minimum value - because of the added dither - so no problem there anymore.
3.Potentially illegal and out of range end point signal levels above max +/-1.0 =0dBfs will clip on a fixed point target format (CD, AES, 24bit files etc.)
Butwhen feeding a floating target (internal DAW etc.) the signal will continue topass without actually clipping - but with quantisation distortion because above +/-1.0 = 0dBfs the dither no longer matches the scaling caused by the changing float exponent..
So in a float system the headroom created by not clipping above max legal signal comes at a cost.
But the saving grace of the entirely circumstantial float headroom is that the quantisation distortion only occurs at levels above flat out which are veryloud anyway - so the chances of hearing the distortion are vanishingly small:-)
--------------------------------------------------------------------------------------------------------------------------------------------------------
JimJohnston told me that if people are monitoring very loud, they won't be able tohear clipping distortion!
Paul :
From experience I would say that's true. Atvery loud levels the ears actually compress too - and the whole can turn into amush that overwhelms our ability to discern much at all, the HF distortioncaused by clipping just joins the other top end mush we hear...
Andin fact creating this 'mush' at lower levels is actually a large part of whatpasses for production these days - because it can sound 'loud and powerful'even when played at lower levels through a boom box, because it has the 'cuesfor loudness' even though it's quiet. I would say the loudness wars are beingdriven largely by this effect. And of course as designers we have made use ofthis kind of thing in real products :-)
Whenit comes to distortion in general though, I am puzzled by it because peopleseem to vary wildly about how much there can be before they perceive it orcomplain about it. Most modern music releases have far more distortion than Ican listen to - but they are selling to others that are not complaining soloudly - so is it just me I often ask?
Butthen again - are people who tolerate it just conditioned to do so, having neverheard much else? DO we become accustomed to it I wonder? Is it even a form ofdeafness that comes from listening to highly compressed music - even if atlower levels? The ear/brain system is highly complex and adaptive - it could becompensating in some way?
Butone thing I can say from research is that it's amazing how much of certainkinds of distortion can become assimilated into the ear/brain system and thenignored if it's part of some recognisable instrument - that is revealed asawful when attached to a sine wave.
Mytheory is that the ear/nervous system is actually highly non-linear and quitevariable, but the brain cleverly corrects this all the time, allowing us tohear sine waves as pure despite the non-linear biological system. Theperception of a pure sine wave is therefore a special case for the ear/brainsystem, which is more easily disturbed than other circumstances..
Ifyou present this system with distortions of the same kind as it's own inherentones, to a large degree it will correct our perceptions of those too - up to apoint..
Atthe point where these distortions are so bad that the ear/brain system isalmost overwhelmed, if the sound is of a previously known and recognisedinstrument (or something else familiar) it helps the brain to deliver arelatively clean perception of it despite the distortion. And in fact it may bethe case that the system can actually learn to do so more effectively over time- if such distortions are present for long periods on program that is highlyfamiliar, like music.
Inresearch I have done I have been amazed to the degree our ear/brain systems areable to adapt - the adaptation has often been so deep that the removal of astimulus can actually make one feel dizzy, disorientated and quite sick!Incredible stuff I would never have imagined before it happened.
----------------------------------------------------------------------------------------------------------------------------------------------------------
now it makes sense to me.
we can only listen to floating point audiodata thru the fixed point window of a D/A converter... and the dither istailored to this D/A....
Paul :
Not just D/A converter - but the real worldin it's entirety!!
Yes- this is the point. Regardless of the math representation within theapplication itself all signal must re-enter the real world in a limited andbounded range.
Allwe are doing is setting what this range is - so that the end product is suitablydithered - so it has no discernible (or even measurable) quantisation artefact- when represented in it's intended output format.
Itisn't clever particularly at all - to me it's pure common sense?
Thefloating representation is simply equivalent to scientific notation - i.e. anumber of significant figures with a multiplier for the range - i.e. 1.234 x10^3 = 1234.???????????. What this tells us is that however large or small theactual value is - it will only ever be represented by 4 digits. These 4 digitscan be ones, tens, hundreds or thousands. What we have is 4 digits of accuracyonly.
In32bit audio we have (just a little less than) 24bits of reliable accuracy - wedecide where in the scale these will be placed.
Andwe have no choice but to align these with real world signal - which is ofcourse fixed point +/-1.0
Sowe decide the lowest level for this signal wrt what is the highest real worldlegal value - and give it the required statistical uncertainty (i.e. dither -noise floor).
Nowinside the processing we can do whatever we like to make use of the floatingsystem (coeffs bigger than +/-1.0 - values smaller than -140dB or so), but whatemerges in the end must conform to the signal requirements of the outputsystem..
Ifyou chose to ignore these requirements by simply 'going for it regardless' infloat applications then there is a penalty - however luckily n most cases it issmall and only happens on very loud signals that effectively mask the errors..
Where it really breaks down is when anapplication is doing something non-linear on purpose - this non-linearity willhave to be referred to some real world signal level - and that is where nastythings can arise if you ignore the actual signal levels involved..
Thisis what happens if you blast extremely high levels through the DSM (apparentlywithout problem) - and then hit the limiter button.
[ 本帖最后由 himhui 于 12-2-18 14:04 编辑 ]
观众反应
最后一个看不下去了，已经眼花了，先这样吧
Cameron:
很好的帖。我坚信对于有关音质的问题，我们如何处理headroom是最重要的影响因素。
有些人认为DAW能处理32 bit浮点是值得讨论的。我不知道多少人注意到这件事：我们在一个32 bit 浮点的DAW里面把音量开大到能听见明显的失真，然后将失真的声音导出为32 bit 浮点的文件.然后把这个文件用wavelab、nuendo、sequoia之类的打开，然后减小波形振幅，让他在0dBFS以下，然后失真就听不到了。当我第一次在nuendo的说明书上读到这个我根本不信直到我试了一遍。我以前以为这样是在毁坏我的母带，后来我认为这对nuendo都不是个挑战，对于neve和api就更不用说了。（最后一句原文似乎很难懂，不知对不对。It has been years since I felt like I waskilling my stereo master and I I don't consider it anymore a challenge inNuendo than it is on our Neve 88R or API.）
Paul的回复：
你说的对。如果所有的处理和信号流程都是严格控制在浮点DAW以内的话，他会对过载有更大的宽容度，你听起来过载了实际上并没有过载，当你减小增益，失真就会消失。但是只有当没有做任何的输出之前才行。（也就说一旦输出当然就变成定点数据了）如果你非常细致的控制整个信号链，让它都是浮点运算的，（虽然很困难）那么你就有更大的headroom和自由。
这种情况主要出现在你使用了第三方的效果卡内部是24bit的精度，PT RTAS的下一个插件如果是TDM，那么数据自然就回不去了。
还有像一些依靠信号本身的电平来反应的处理，例如压缩，限制，最大化。还有一些加味道的效果器，使用了一些非线性的处理来改变泛音特点的。这些效果器处理大0dBFS的信号一般都会不怎么好（有些甚至很糟）。
DSM（paul自己研发的效果器）是能够处理0dBfs左右的信号的，而且控制的范围也很大，阈值最小有-24 dBfs，输出增益可达到24 dBfs。
它内部也是浮点运算，并且在浮点允许的范围内不会有失真。
当然他还有限制器部分，限制器就像别的标准限制器一样，最大允许的电平是0 dBfs（这是输出介质决定的），所以信号到了限制器部分就会回到我们真实的电平。
限制器部分最高能支持+6dBfs的信号（最后输出的是0dBfs），将信号推到+6还是可以的，超过+6就会削波。
Astimony:
没有人会质疑您在这个领域的专业地位。但是我读过一份论文，他描述了32
Bit浮点的抖动。他有3个可靠的作者，两个是布达佩斯大学一个是斯坦福大学的。我主要是一个程序员，但是关于那份论文中说的32bit的抖动应用我还是有很大的困难。
您能否共享一些类似的代码？不是真的也可以，只要能看出如何抖动32bit的浮点数据，或者真的也可以？
或者我猜你是先把数据转换成定点的处理完再抖动/处理成32bit的？我根据您以前的发言猜的。
如果你能告诉我就太好了，或许就能解决抖动量化错误带来失真的问题。
当然如果不能分享也能理解，毕竟您有商业上的考虑。
Paul的回复：
这篇论文没有错，一般来说是这样。
但是我们的目的是在可能的范围内去除泛音失真。这个范围是存在的，就是底噪音量到设备0点之间的范围，不管他是用什么方法去表达出来的，浮点还是定点。那么不管什么方法，都要能完整表达这大约140dB内的所有值，还要能用抖动来消除量化失真。
举个例子，DSM是浮点运算（内部64bit双倍精度），我们在输出还是要把数据抖动到我们需要的范围内，例如-140 —— 0 dB。也就是说输入的是32bit浮点还是24bit定点都不能有失真。
换句话说，无论信号在数学表达上是什么样的，我们要让他变成真正的电流，并且要在140dB的范围内。
那么实践中会怎么样？
1如果插件输出了一个已知信噪比的浮点信号（在浮点运算时已经抖动过），那么转换成32bit浮点或者24bit定点是都不会有量化失真。结果会怎样并不取决于我们，而是DAW。有些能处理64bit浮点，有些是32bit，像TDM卡，powercore卡只能处理24bit定点。所有最终输出其实都是24bit定点。如果你需要16bit，DSM也提供了抖动的开关。
2 信号电平比底噪还要小的话是无法取值的，所以没问题。
3 最大值超出范围，在定点的格式上会造成失真（CD,AES,24bit文件等）。但是如果输出的下一级处理是浮点的，比如DAW内部数据，信号会继续传输不会过载，但是会带来bit转变时的量化错误，带来失真。
所以综上，浮点的系统高headroom是有代价的。
但是，这样的量化错误仅仅发生在信号值超出浮点的最大值的时候，实际上这时已经极为大声，所以听到这样的失真几率基本等于零。
（这一段涉及到一些基础原理，我翻译不是很有把握，看不懂还是以原文为准。）
JimJohnston告诉我如果监听音量非常大，那么就听不到削波失真了？
Paul的回复：我个人的经验认为这是对的。音量很大的时候我们的耳朵就会压缩音量，这时候声音就被压成一团浆糊，很多细节都听不到了。
事实上把声音压成一团是今天制作的很大一部分，因为就算你在低音量下播放，那样强烈压缩过的音乐会使你感觉仿佛他在大声播放一样，就算小音量在一般的烂音箱放，他也给人响而有力的感觉，因为他在用强烈的压缩暗示大脑。响度战争很大程度上就是因此而起。然而我们软件设计者也会想利用它。
当声音失真的时候，我很困惑，每个人表现的很不一样。大部分现代音乐对于我来说失真都太大了，但是买它的那些人却没有什么抱怨。我经常觉得是不是只有我特殊一点？还有，那些能容忍失真的人是适应了，还是从来没听过不失真的音乐？我们是不是习惯了这样？是不是我们已经产生了一些听觉缺陷尤其是对于高度压缩的音乐—就算这个音乐播放的音量不大？大脑是不是对音乐做出了某些补偿？
有一点可以确定就是研究显示有些特定的乐器的失真会被接受，但是如果这种失真用在正弦波上就会被所有人认为是难听。我的想法就是大脑在随时的补偿和改变我们听到的内容，心理因素也会影响听到的。（此处省略一些）我很吃惊大脑能渐渐适应失真，甚至我把一些失真去除以后音乐反而变乏味了，这是以前没想到的。
很好的帖。我坚信对于有关音质的问题，我们如何处理headroom是最重要的影响因素。
有些人认为DAW能处理32 bit浮点是值得讨论的。我不知道多少人注意到这件事：我们在一个32 bit 浮点的DAW里面把音量开大到能听见明显的失真，然后将失真的声音导出为32 bit 浮点的文件.然后把这个文件用wavelab、nuendo、sequoia之类的打开，然后减小波形振幅，让他在0dBFS以下，然后失真就听不到了。当我第一次在nuendo的说明书上读到这个我根本不信直到我试了一遍。我以前以为这样是在毁坏我的母带，后来我认为这对nuendo都不是个挑战，对于neve和api就更不用说了。（最后一句原文似乎很难懂，不知对不对。It has been years since I felt like I waskilling my stereo master and I I don't consider it anymore a challenge inNuendo than it is on our Neve 88R or API.）
Paul的回复：
你说的对。如果所有的处理和信号流程都是严格控制在浮点DAW以内的话，他会对过载有更大的宽容度，你听起来过载了实际上并没有过载，当你减小增益，失真就会消失。但是只有当没有做任何的输出之前才行。（也就说一旦输出当然就变成定点数据了）如果你非常细致的控制整个信号链，让它都是浮点运算的，（虽然很困难）那么你就有更大的headroom和自由。
这种情况主要出现在你使用了第三方的效果卡内部是24bit的精度，PT RTAS的下一个插件如果是TDM，那么数据自然就回不去了。
还有像一些依靠信号本身的电平来反应的处理，例如压缩，限制，最大化。还有一些加味道的效果器，使用了一些非线性的处理来改变泛音特点的。这些效果器处理大0dBFS的信号一般都会不怎么好（有些甚至很糟）。
DSM（paul自己研发的效果器）是能够处理0dBfs左右的信号的，而且控制的范围也很大，阈值最小有-24 dBfs，输出增益可达到24 dBfs。
它内部也是浮点运算，并且在浮点允许的范围内不会有失真。
当然他还有限制器部分，限制器就像别的标准限制器一样，最大允许的电平是0 dBfs（这是输出介质决定的），所以信号到了限制器部分就会回到我们真实的电平。
限制器部分最高能支持+6dBfs的信号（最后输出的是0dBfs），将信号推到+6还是可以的，超过+6就会削波。
Astimony:
没有人会质疑您在这个领域的专业地位。但是我读过一份论文，他描述了32
Bit浮点的抖动。他有3个可靠的作者，两个是布达佩斯大学一个是斯坦福大学的。我主要是一个程序员，但是关于那份论文中说的32bit的抖动应用我还是有很大的困难。
您能否共享一些类似的代码？不是真的也可以，只要能看出如何抖动32bit的浮点数据，或者真的也可以？
或者我猜你是先把数据转换成定点的处理完再抖动/处理成32bit的？我根据您以前的发言猜的。
如果你能告诉我就太好了，或许就能解决抖动量化错误带来失真的问题。
当然如果不能分享也能理解，毕竟您有商业上的考虑。
Paul的回复：
这篇论文没有错，一般来说是这样。
但是我们的目的是在可能的范围内去除泛音失真。这个范围是存在的，就是底噪音量到设备0点之间的范围，不管他是用什么方法去表达出来的，浮点还是定点。那么不管什么方法，都要能完整表达这大约140dB内的所有值，还要能用抖动来消除量化失真。
举个例子，DSM是浮点运算（内部64bit双倍精度），我们在输出还是要把数据抖动到我们需要的范围内，例如-140 —— 0 dB。也就是说输入的是32bit浮点还是24bit定点都不能有失真。
换句话说，无论信号在数学表达上是什么样的，我们要让他变成真正的电流，并且要在140dB的范围内。
那么实践中会怎么样？
1如果插件输出了一个已知信噪比的浮点信号（在浮点运算时已经抖动过），那么转换成32bit浮点或者24bit定点是都不会有量化失真。结果会怎样并不取决于我们，而是DAW。有些能处理64bit浮点，有些是32bit，像TDM卡，powercore卡只能处理24bit定点。所有最终输出其实都是24bit定点。如果你需要16bit，DSM也提供了抖动的开关。
2 信号电平比底噪还要小的话是无法取值的，所以没问题。
3 最大值超出范围，在定点的格式上会造成失真（CD,AES,24bit文件等）。但是如果输出的下一级处理是浮点的，比如DAW内部数据，信号会继续传输不会过载，但是会带来bit转变时的量化错误，带来失真。
所以综上，浮点的系统高headroom是有代价的。
但是，这样的量化错误仅仅发生在信号值超出浮点的最大值的时候，实际上这时已经极为大声，所以听到这样的失真几率基本等于零。
（这一段涉及到一些基础原理，我翻译不是很有把握，看不懂还是以原文为准。）
JimJohnston告诉我如果监听音量非常大，那么就听不到削波失真了？
Paul的回复：我个人的经验认为这是对的。音量很大的时候我们的耳朵就会压缩音量，这时候声音就被压成一团浆糊，很多细节都听不到了。
事实上把声音压成一团是今天制作的很大一部分，因为就算你在低音量下播放，那样强烈压缩过的音乐会使你感觉仿佛他在大声播放一样，就算小音量在一般的烂音箱放，他也给人响而有力的感觉，因为他在用强烈的压缩暗示大脑。响度战争很大程度上就是因此而起。然而我们软件设计者也会想利用它。
当声音失真的时候，我很困惑，每个人表现的很不一样。大部分现代音乐对于我来说失真都太大了，但是买它的那些人却没有什么抱怨。我经常觉得是不是只有我特殊一点？还有，那些能容忍失真的人是适应了，还是从来没听过不失真的音乐？我们是不是习惯了这样？是不是我们已经产生了一些听觉缺陷尤其是对于高度压缩的音乐—就算这个音乐播放的音量不大？大脑是不是对音乐做出了某些补偿？
有一点可以确定就是研究显示有些特定的乐器的失真会被接受，但是如果这种失真用在正弦波上就会被所有人认为是难听。我的想法就是大脑在随时的补偿和改变我们听到的内容，心理因素也会影响听到的。（此处省略一些）我很吃惊大脑能渐渐适应失真，甚至我把一些失真去除以后音乐反而变乏味了，这是以前没想到的。
观众反应
