Dithering to 16-bit Quandry?

[Dave Labrecque]

I'm outputting the final 24-bit mix for a film. The editor also wants a 16-bit version from me. I put the 24-bit mix on the timeline in the MT in a virgin session, and want to put a Levelizer on the final output, limiter set to the normalized max of the 24-bit file (97%), thinking that the dither I'm adding while building the 16-bit file could push some peaks over that 97% limit that I want maintained if I don't brickwall limit it.

The help file seems to indicate that the dither is added (I'm doing it in the buildmix dialog; not sure if the injection point is different than that for the MT dither) just prior to the Final Res patch point. Whoa -- never knew that before. So, that would be the logical point to add the final brickwall, but the Final Res patch point won't allow buffer-size-altering plugs, which the Levelizer is when it's limiting.

So... how does one keep SAW's native dither from pushing peaks over the desired ceiling? I've been dithering for years and not even stopping to think about this. Many of my 16-bit masters have apparently been going out with peaks over the ceiling I'd thought I was getting.

[Bob L]

You use a lower final level... like 95%... you can also look at the final 16 bit result and go back and use a lower limit if the signal still peaks over after the dither.

You have to use your best guess when dither is involved.

Or... you can add the dither without limiting to the max, then go back and brickwall the final after the dither has been added.

Bob L

[Dave Labrecque]

You use a lower final level... like 95%... you can also look at the final 16 bit result and go back and use a lower limit if the signal still peaks over after the dither.

You have to use your best guess when dither is involved.

Or... you can add the dither without limiting to the max, then go back and brickwall the final after the dither has been added.

Bob L

Of course... brickwall/normalize it on a second pass. That'll do it. Thanks.

[DominicPerry]

Of course... brickwall/normalize it on a second pass. That'll do it. Thanks.

Sure your client won't hear the lack of dither on those limited peaks? ;)

Dominic

[DominicPerry]

Of course... brickwall/normalize it on a second pass. That'll do it. Thanks.

I think you want to avoid normalizing, just peak limit. Normalizing will undo all of your dither. I've just been reading an earlier thread about dithering, and I'm very confused - you'd better check.

Dominic

[Cary B. Cornett]

First of all, I'm with Bob's suggestion. I would think that limiting to 95% would leave more than enough headroom for any peaks added by the dither. As I understand it, the peak amplitude of the dither is normally equal to the level of the Least Significant Bit at the output resolution, and at 16 bits that's mighty tiny! Any added peaks would, I think, be very small indeed.

[Bob L]

Normalizing with the Levelizer on a second pass would not affect the dither data to any real degree.... It may raise the noise floor a bit... but the normalize factor on the lowest bits would become fairly insignificant.

Bob L

[Dave Labrecque]

Sure your client won't hear the lack of dither on those limited peaks? ;)

Dominic

The dither should still be there. And I'm sure they won't hear it. :)

[Dave Labrecque]

Normalizing with the Levelizer on a second pass would not affect the dither data to any real degree.... It may raise the noise floor a bit... but the normalize factor on the lowest bits would become fairly insignificant.

Bob L

FWIW... I checked the limited peaks to which dither was added (Levelizer set to limit at 97% and patched on the master output, post-fader, dither added during buildmix) and they still don't exceed 97.00%. This would seem to imply that the dither (Bob's Type 1) is less than 0.01% in level. Could it be?

[Cary B. Cornett]

FWIW... I checked the limited peaks to which dither was added (Levelizer set to limit at 97% and patched on the master output, post-fader, dither added during buildmix) and they still don't exceed 97.00%. This would seem to imply that the dither (Bob's Type 1) is less than 0.01% in level. Could it be? Yes it could. Again, the added dither signal has a peak level about equal to the Least Significant Bit at the output resolution. IIRC, for 16 bits that would be 1 out of 16,384... So yeah, a little under 0.01%.

I think the confusion here may be about what can happen with sample rate conversion, which is NOT the same as dithering down to fewer bits per sample. In SRC there is usually some kind of anti-aliasing filter involved, and that filtering can introduce some degree of overshoot, forcing you to lower your maximum peak level to allow "headroom" for the overshoot. How bad the overshoot is depends on the design/coding of the filter.

The only time when you should have to worry about peak overshoot when dithering is when you are ALSO applying SRC.

[Dave Labrecque]

Yes it could. Again, the added dither signal has a peak level about equal to the Least Significant Bit at the output resolution. IIRC, for 16 bits that would be 1 out of 16,384... So yeah, a little under 0.01%.

I think the confusion here may be about what can happen with sample rate conversion, which is NOT the same as dithering down to fewer bits per sample. In SRC there is usually some kind of anti-aliasing filter involved, and that filtering can introduce some degree of overshoot, forcing you to lower your maximum peak level to allow "headroom" for the overshoot. How bad the overshoot is depends on the design/coding of the filter.

The only time when you should have to worry about peak overshoot when dithering is when you are ALSO applying SRC.

Is it an accepted approach, then, to normalize as needed, and simply not worry about the small increase in level that dither will incur?

Steve Berson? I think you've said you normalize to 97%. Do you add dither "on top of" that?

[TotalSonic]

Is it an accepted approach, then, to normalize as needed, and simply not worry about the small increase in level that dither will incur?

Dither should be added as the very final step (after any additional processing, including any peak limiting, gain staging such as normalizing, or sample rate conversion) only whenever (and at the same time) that requantizing occurs.


Steve Berson? I think you've said you normalize to 97***37;. Do you add dither "on top of" that?

I dither as the final step AFTER any limiting or processing!! Adding dither does not function to correctly smooth out distortions caused by requantization if it is not the very final process!!

I generally leave a -0.3dBFs output ceiling - of which the closest setting for the Levelizer in percentage is 97%. For algorithms that do heavy clipping (i.e. GVST GClip) that have output ceiling labelled in percentage I find sometimes you have to have output ceiling set even lower (generally around 96%) if you want to avoid intersample peak overs. The Sonoris Meter is one tool that can check this for you (if you place it in oversampling mode).

A good read on the subject of dither is at http://www.cadenzarecording.com/dither.html

Best regards,
Steve Berson

[Dave Labrecque]

We seem to be miscommunicating. It sounds like you DO add dither after normalizing, which will increase the signal over 97%.

Is this correct?

[TotalSonic]

We seem to be miscommunicating. It sounds like you DO add dither after normalizing, which will increase the signal over 97***37;.

Is this correct?

Yes - if you are requantizing (going from 24bit to 16bit) then any dither added should be added as the last process. If your Levelizer is set to normalize to 97% then in the vast majority of cases there should be more than enough headroom to accomodate the additional overall gain added by the dither placed after the Levelizer without risking any overs. With some material that has been heavily clipped then there is indeed the possibility that you need to set the output ceiling lower though, if you wish to avoid any intersample peak overs.

Best regards,
Steve Berson

[Dave Labrecque]

Yes - if you are requantizing (going from 24bit to 16bit) then any dither added should be added as the last process. If your Levelizer is set to normalize to 97% then in the vast majority of cases there should be more than enough headroom to accomodate the additional overall gain added by the dither placed after the Levelizer without risking any overs. With some material that has been heavily clipped then there is indeed the possibility that you need to set the output ceiling lower though, if you wish to avoid any intersample peak overs.

Best regards,
Steve Berson

Thanks, Steve. How much of an intersample peak over do you consider to be too much? Insofar as the dither is already (slightly) over (in this case) 97%, how much more is too much, IYHO?

[TotalSonic]

Thanks, Steve. How much of an intersample peak over do you consider to be too much? Insofar as the dither is already (slightly) over (in this case) 97***37;, how much more is too much, IYHO?

The exact amount of headroom isn't the important thing - you only have to worry about whether any peak in the file actually goes over 0dBFs! - that's why allowing for some headroom after limiting to accomodate the additional gain the dither noise is adding is recommended.

Leaving about 0.1 - 0.5dBFs worth of headroom after limiting is also a good idea to allow a little bit of headroom so that intersample peak overs don't occur (which when they do can lead to a distorted output on some cheaper consumer systems) - and also so after any reprocessing such as mp3 encoding, or consumer end digital processors such as iTunes "Sound Enhancer" that the new re-encoded files aren't distorted.

Again at what point intersample peak overs occurs when the output is pushed closer to 0dBFs varies from file to file - i.e. heavily clipped files will need more headroom below 0dBFs - while natural wave forms generally can come up to 0dBFs without a problem. A small amount of intersample peak overs occuring in playback certainly won't "ruin" a file in most instances - but for cleanest output on all systems I recommend avoiding them. A lot of times when clients request ultra-slammed levels its an issue that's impossible to avoid though. For what you are describing I can't imagine the extra headroom causing any problems for the end client though.

If you're worried about clipping overs or intersample peak overs then I also recommend getting something like the Sonoris Meter which allows you to check for them.

Otherwise - approximately 0.3dBFs output ceiling (i.e. setting the Levelizer's "normalize" setting to 97% maximum) should be fine general rule that allows you to avoid the various problems (which generally are minor) encountered with hitting 0dBFs in the vast majority of cases.

Best regards,
Steve Berson

[Dave Labrecque]

The exact amount of headroom isn't the important thing - you only have to worry about whether any peak in the file actually goes over 0dBFs! - that's why allowing for some headroom after limiting to accomodate the additional gain the dither noise is adding is recommended.

Leaving about 0.1 - 0.5dBFs worth of headroom after limiting is also a good idea to allow a little bit of headroom so that intersample peak overs don't occur (which when they do can lead to a distorted output on some cheaper consumer systems) - and also so after any reprocessing such as mp3 encoding, or consumer end digital processors such as iTunes "Sound Enhancer" that the new re-encoded files aren't distorted.

Again at what point intersample peak overs occurs when the output is pushed closer to 0dBFs varies from file to file - i.e. heavily clipped files will need more headroom below 0dBFs - while natural wave forms generally can come up to 0dBFs without a problem. A small amount of intersample peak overs occuring in playback certainly won't "ruin" a file in most instances - but for cleanest output on all systems I recommend avoiding them. A lot of times when clients request ultra-slammed levels its an issue that's impossible to avoid though. For what you are describing I can't imagine the extra headroom causing any problems for the end client though.

If you're worried about clipping overs or intersample peak overs then I also recommend getting something like the Sonoris Meter which allows you to check for them.

Otherwise - approximately 0.3dBFs output ceiling (i.e. setting the Levelizer's "normalize" setting to 97% maximum) should be fine general rule that allows you to avoid the various problems (which generally are minor) encountered with hitting 0dBFs in the vast majority of cases.

Best regards,
Steve Berson

Wow. The more I learn, the less, it turns out, I know. :o I found this, which reinforces what you say, and takes it a little further:

http://www.gearslutz.com/board/tips-techniques/334385-intersample-peaks.html

Looks like the bottom line is that there's no well-defined bottom line. As long as we're trying to get stuff close to the top of the scale, we're going to be risking artifacts on some system somewhere. And it's kind of a (n educated) guessing game as to what you can get away with based on type of music, how much mastering processing, what the destination lossy encoding possibilities are and what quality of DAC's will be on the other end.

No fun. :(

I guess that's why we should hire guys like you, Steve. How much do you charge to master a :30 radio spot? :)

[Grekim]

Relax about the dither. If in doubt, just playback your 16 bit master and note the peaks. Not a bad idea anyway for other reasons.

You can really get into trouble with sample rate converting a track that is at or near peak. Actually, I can't speak about SAW directly, but it has been my experience in every other DAW that you can have as much 1.2 dB difference, which would be disasterous if you're already at -0.3 dBFS before converting.

[DominicPerry]

Not sure if this helps, but I just ran the Levelizer with the Sonoris meter and can get an OS peak at 93% normalizing if you peak limit at 5%. I know this is silly, but I can get a OS peak at 97% with peak limiting at 40% which has no audible artifacts on the music I'm doing this to. This is all before dither. I'm sure with hours of experiment I can discover 'safe' and 'unsafe' numerical limits on the use of peak limiting, normalizing and dither to avoid OS peaks. The question still seems to be about what you can and can't hear. 5% normalizing is vile, regardless of OS peaks. What does the occasional OS peak sound like on one of those CD players which can't cope with it? Probably inaudible - it's more likely that the listening pleasure is ruined by a thoroughly crushed mix rather than a hint of distortion for a millisecond. Dunno. But not worth getting obsessed with dither itself, I suspect.

Dominic

PS, I can't really hear dither myself. Sorry.

[Dave Labrecque]

Not sure if this helps, but I just ran the Levelizer with the Sonoris meter and can get an OS peak at 93% normalizing if you peak limit at 5%. I know this is silly, but I can get a OS peak at 97% with peak limiting at 40% which has no audible artifacts on the music I'm doing this to. This is all before dither. I'm sure with hours of experiment I can discover 'safe' and 'unsafe' numerical limits on the use of peak limiting, normalizing and dither to avoid OS peaks. The question still seems to be about what you can and can't hear. 5% normalizing is vile, regardless of OS peaks. What does the occasional OS peak sound like on one of those CD players which can't cope with it? Probably inaudible - it's more likely that the listening pleasure is ruined by a thoroughly crushed mix rather than a hint of distortion for a millisecond. Dunno. But not worth getting obsessed with dither itself, I suspect.

Dominic

PS, I can't really hear dither myself. Sorry.

What's an OS peak?

[TotalSonic]

What's an OS peak?

When you place the Sonoris Meter in "OS" (Over Sampled) mode it can detect intersample peak overs.

A few of these occurring every so often certainly is not going to create a problem in the vast majority of cases. If you see these happening constantly then it might be a good idea to give things a tiny bit more headroom though.

I should note that there are tons of examples of current commercial releases that either go all the way to 0dBFs and/or have a significant amount of intersample peak overs. Of course lots of releases these days have degraded sound so this fact is in no way a recommendation to do the same!

btw - if you're indeed ever interested in having me master anything for you then just PM at steve at totalsonic dot net and I can answer any questions and set your order up. My basic rates are up at http://www.totalsonic.net/rates.htm - but anyone on this board qualifies for my "friends and family" discount off my already incredibly low rates.

Best regards,
Steve Berson

[Sean McCoy]

I remember hearing somewhere that some or all mastering houses would reject any CD master that contained even a single over. Not true?

[TotalSonic]

I remember hearing somewhere that some or all mastering houses would reject any CD master that contained even a single over. Not true?

Depends on what you mean by "mastering houses". If you are referring to facilities for the glass mastering process necessary to create stampers for making replicated pressed CD's then the answer is yes - in the early days of CD production if an over was detected then they would reject the master as out of spec. This practice, however, was abandoned by the early 90's - and at this point I know of no replication plant that rejects provided masters for containing overs.

If you are referring to current audio mastering studios then the answer is no - I unfortunately receive mixes all the time that have some peaks clipped or that have been brickwall limited already. I always ask first for a version that has some headroom and/or without any peak limiting placed on the 2-bus - but sometimes clients are not able to provide new versions - in which case I attenuate the source prior to any processing and do my best with what has been given to me.

Best regards,
Steve Berson

[Dave Labrecque]

Interesting... I downloaded Stillwell's "Bitter" and SSL's "X-ism", both free plugs that are supposed to show clips, IS overs, and bits used. And they both give completely different results! Bitter is really hard to get to show IS overs. While X-ism is pretty happy to "reveal" them. Obviously they're using different reconstruction algos. Who's a guy to believe? :confused: :p :( :rolleyes:

I thought I read somewhere that RME's Digicheck has an IS indicator in there somewhere, but I can't find it.

[Dave Labrecque]

The most stable release of any operating system.

Heh, heh, heh...

[TotalSonic]

Interesting... I downloaded Stillwell's "Bitter" and SSL's "X-ism", both free plugs that are supposed to show clips, IS overs, and bits used. And they both give completely different results! Bitter is really hard to get to show IS overs. While X-ism is pretty happy to "reveal" them. Obviously they're using different reconstruction algos. Who's a guy to believe? :confused: :p :( :rolleyes:

I thought I read somewhere that RME's Digicheck has an IS indicator in there somewhere, but I can't find it.

Dave -
The problem with getting overly concerned with this issue is that it is a moving target - in the same way that every meter that uses a different algorithm show different results - every playback system will be effected a little differently - with some DAC's not being bothered in the least - to some others creating audible crackles at the points where the overs are particularly bad.

Please realize though that in most cases this is a non-issue - just keep a tiny bit of headroom and you'll find no one should have issues. If you're really trying to keep things ultra-clean then don't clip the audio at any point in your chain and use -0.5dBFs as an output ceiling. Otherwise anywhere around a -0.1dBFs to -0.3dBFs output ceiling should work fine in the majority of cases.

Best regards,
Steve Berson

[Dave Labrecque]

Dave -
The problem with getting overly concerned with this issue is that it is a moving target - in the same way that every meter that uses a different algorithm show different results - every playback system will be effected a little differently - with some DAC's not being bothered in the least - to some others creating audible crackles at the points where the overs are particularly bad.

Please realize though that in most cases this is a non-issue - just keep a tiny bit of headroom and you'll find no one should have issues. If you're really trying to keep things ultra-clean then don't clip the audio at any point in your chain and use -0.5dBFs as an output ceiling. Otherwise anywhere around a -0.1dBFs to -0.3dBFs output ceiling should work fine in the majority of cases.

Best regards,
Steve Berson

OK, thanks, Steve.

[TotalSonic]

Speaking of meters ....I have only a couple of days to add the Dorrough Meter bundles to my Waves "NARAS discount", or wait another year.

I've been delayed updating my OS, etc, due to a defective "boot drive" backup.

But, any opinions on the dorrough software meters. (not even sure if they are available VST or DX

mark

I haven't seen the Waves version except for a YouTube promo on them - but I'm familiar with hardware Dorrough's - imho they are about the best out there in combined PPM/VU LED meters - although for my own studio I got much cheaper and more compact Logitek meters as they offer the same functionality (if on my particular unit not the same resolution) as the Dorroughs.

As far as the Waves version of the Dorrough meters - I think their ability to resize is a pretty cool feature - but for much less money you actually get more functionality with the Sonoris Meter 2 - in that the Sonoris Meter has options for K-System metering and for measuring average loudness not only as RMS but also as LEQ-A - and which is now available as VST for non-SAW apps (although not yet ported over to Mac) http://www.sonoris.nl/catalog/meter-p-34.html

Otherwise for less money than the Waves Dorrough you can source a hardware Logitek unit and be able to monitor analog and digital sources without having to even have your DAW turned on - and also never worry about authorizations, WUP fees, or compatibilities - http://www.logitekaudio.com/products.html

Best regards,
Steve Berson

[TotalSonic]

Checking street price on the Waves Dorrough seems it's going for $375 which is a bit less than I thought previously (although still a good bit more than the $125 that Sonoris is asking for their meter). So if the NARAS deal is for half the direct price of $500 - making it $250 - then it still is cheaper than hardware.

Anyway doing a search on ebay actually turns up a few used Dorroughs for reasonable amounts of cash. To me if you're going to bother to pony some actual cash for a meter it still makes sense to get the hardware version.

Best regards,
Steve Berson

[TotalSonic]

Steve, I'm not up to speed on the different Dorrough hardware options, although I was quite impressed with the Waves Dorrough demo at NAMM.

I found this on ebay. Is this comparable?

http://cgi.ebay.com/Dorrough-40-C2-Analog-Audio-LED-PPM-Peak-RMS-VU-Meter_W0QQitemZ250351781551QQcmdZViewItemQQptZLH_D efaultDomain_0?hash=item250351781551&_trksid=p3286.c0.m14&_trkparms=72%3A1234%7C66%3A2%7C65%3A12%7C39%3A1%7C 240%3A1318%7C301%3A1%7C293%3A1%7C294%3A50

Looking at the Waves site seems the bundle gives you emulations of the 280D/240D, 380D/340D, and 40AES/EBU. I believe this model would give you the same as the 40AES/EBU emulation except that this is setup with analog input.

If you go the hardware rout you need to what best fits your already existing setup. This might mean that either you have an existing analog feed in your console or monitor controller in which case get a meter with analog input - or extra DAC output that you can mult the signal you want to monitor with (i.e. if you have a soundcard with say 8 channels of output and you only use 2 of them then ) in which case analog inputs would work as well. Otherwise if you have an extra AES or spdif send from your DAW, digital console, or digital patchbay, then a meter with an AES input could work better for you.

My Logitek meter is 6 channel with analog ins - so I send the meter output from my Coleman Audio monitor controller for first 2 channels so I can always get a read of what I'm sending out to my monitors.

I also have an extra output from my Apogee 896i/o (which is connected to my Sydec Mixtreme 192 for both digital and extra analog i/o to my main DAW) so I can monitor the final output level on my main DAW - nice to have as I never need to bring a software meter to the top of my DAW's screen if want to check my peak and VU levels.

For the last set of meters I have a send from my phono pre-amp's auxilliary output's - nice for checking levels on test records or doing other such comparisons (doesn't happen that often - but it's nice as to have as a service for clients who still get vinyl cut from pre-masters I make for them).

Although kind of a "luxury" in this day when so many things in the studio these days are cheaper and easier to get virtual versions of - I still think hardware meters have a few advantages.

Best regards,
Steve Berson