Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Dave Scoven]

I'm sure this topic's been done before, but...

Can someone please explain exactly what the theory is behind mixing a project tracked at 24/48 "up" to 88 or 96k (then back to 44.1)?  You see the advantage asserted all the time, but is there any real theoretical (not to mention audible) evidence to support the assertions?  It just seems to me that tracks recorded at 24/48 will not gain anything by being mixed at a higher sample rate (unless the idea is that it makes a difference in terms of effects and processing).

Help?  Thanks!

[bruno putzeys]

The topic is bound to have been discussed at length in Dan's former forum, so here's a short run-down:

In favour of higher sampling rates: Some processes will benefit from the higher sampling rates.
*Comp/limit processes will be better able to track the signal amplitude when processing high frequency contents.
*The frequency response from EQ's will better match that of equivalent analog EQ's. At frequencies near the nyquist limit, the frequency response of IIR filters gets aliased (not the signal itself, we're talking about a linear effect here).

Against:
*The extra conversion step is usually done less than picture perfect.
*EQ's will produce more garbage when working at low frequencies.

The best practice is to track and mix at the highest sampling rate you're actually going to release in. When you're sure the project is going to CD only, track and mix at 44.1kHz. Use EQ plugins that correct for response warping like the Oxford plugins. Use comp/limit plugins that have an upsampler in the side chain (e.g. again Oxford) but not in the signal chain.
If you must use a higher sampling rate to accommodate for deficiencies in the EQ and other plug-ins, use an integer multiple (88.2kHz for CD). The most insane thing to do is working at 96kHz for a CD only project.

[Dave Scoven]

Bruno,

VERY insightful and understandable.  Thank you.  Now, what about analog processing?  Suppose you track to HD, but use analog processing to mix?  

Also, I used to instinctively follow the track and mix at 44.1k like of reasoning because I reckoned that requires zero conversion.  Then I tried bumping up to 24/48.  I didn't notice any difference on most things.  Except drums.  For some reason, 24/48 seemed to capture drum sounds with much more accuracy.  At 16/44.1, I was pulling (what's left of) my hair out trying to get the "whole" drum sound, but at 24/48, it all came together.  Or came together with much higher fidelity.  Of course, THAT sparked the "I should be tracking at 24/88.2!!" thing.

Do you think that higher sample rates better capture percussive transients?  

[bruno putzeys]

It's quite clear that 44.1 has always been a bit marginal. As the difference with 48kHz attests, the boundary is somewhere around there.
However, when you're recording for CD you have to keep in mind that all the quality improvement you hear when working at 48k or higher is lost when converting back to 44.1k. You're the only one to enjoy better reproduction. If this goes at the expense of the quality of the end product (e.g. by working at 48kHz and having to sample rate convert afterwards) this is not such a good idea.

I can't really imagine any analogue processing that would work better when inserted in a high-sampling-rate chain than in a 44.1k chain. Apart perhaps from the most nonlinear of processes (a fuzz box) there is nothing about the >20k spectrum that would have much effect on the <20k spectrum. And beyond the spectral content of the input signal, analogue effects devices have no idea about the sampling rate of the system they're sitting in.

I should say that when you're mixing in the analogue domain, you could of course track at any other rate and then take down the mix at 44.1k. In fact, I would make a choice between either running the master recorder and the multitrack on the house sync (and that be 44.1kHz) or running the multitrack at a substantially different sampling frequency (48k or higher). Most AD/DA converters have a region between 0.45fs and 0.55fs (for 44.1k that's between 20.000kHz and 24.100kHz) where aliases and images are left relatively unattenuated. On AB miked material this effect is audible as an obvious smearing of wideband noises (breath, transients) across the whole stereo image.
If the multitrack and the master recorder are running asynchronously but both nominally at the same rate, both will contribute with slightly different products. It's similar to running an asynchronous sample rate converter with nearly equal sample rates, which is also not a good idea.

The alias story also points towards how 44.1k can be improved. To do so with existing replay equipment in mind one needs to add an even steeper lowpass filter right at 20kHz. Inserting a steeper filter cleans up the stereo image spectacularly, although I wouldn't say it helps transient reproduction much. The improvement is also not so pronounced on intensity stereo or "multi mono" mixes.

[Dave Scoven]

Bruno,

First, thank you for the really understandable and information-packed replies.  I'm learning a lot here.

I've noticed that in digital linear phase mode, Ozone's HP & LP filter slopes are REALLY steep.  Would something like that be sufficient to produce the cleanup effect @ 20k?  What about @18k?

[bruno putzeys]

I don't know the Ozone but the filter I use is around 350 taps long. It goes from 0dB to -90dB between 19.75kHz and 20.0kHz (that filter is part of a "any format to any format" box I'm developing, so not yet available).
Arguably something less steep starting earlier may also do, but I've not yet done the test.

(note: some edits in previous post).

[Dave Scoven]

Thanks Bruno.  Just one more question.

I'm in the middle of a project tracking at 24/48.  Would you reccomend
a) Mixing "up" to 88.2k, then back to 44.1, or
b) Mixing at 48k, then back to 44.1?

[bruno putzeys]

So you're running a multitrack at 48k, mix analogue and taking down the mix at 44.1k? Or else, could you describe the setup in a bit more detail?

[Dave Scoven]

Bruno,

Sure.  For my current project I'm multitracking at 24/48 (MOTU 896HD/Digital Performer). I mix in DP, but processing is analog hardware (1176s +  Neve, Focusrite and UA channels for EQ).  I do use Altiverb and Ozone plugins.

After reading your responses, I wish I had tracked from the beginning at 88.2k, but I didn't.  SO -- should I mix at 88.2, then back to 44.1 for cd, or.. should I mix at 48.  Seems like I should mix at 88.2 because that might be better for the plugs, then back to 44.1.

Am I on track?

[bruno putzeys]

That hinges on how good the sample rate conversion is. If it's correctly done, going from 48 to 88.2 and on to 44.1 should produce the same result as going straight from 48 to 44.1. In that case there's little wrong with going through 88.2, except that Altiverb would require nearly 4 times as much processing power at 88.2.

It would appear to me that it's simplest to keep the project in 48k now and convert to 44.1 when you're done.

[Dave Scoven]

I see that.  Thanks so much.  So what I'm taking away from this is
a) that the option to work at higher sampling rates is really there for people whose final product is a medium other than cd, but
b)while working at 24/48 might be slight overkill for cd projects, if it helps me mix because I can hear things more clearly while working at 48, then its probably worth it, even though much of that clarity will be lost in the final product.  Fair?

[bruno putzeys]

For b) I'd suggest 88.2 because of the easier conversion to 44.1.

[Dave Scoven]

Terrific.  Bruno, it's been a pleasure.  Thanks so much for the great advice.  I learned a lot.

Dave

[krabapple]

Dave Scoven wrote on Thu, 18 January 2007 18:38

Also, I used to instinctively follow the track and mix at 44.1k like of reasoning because I reckoned that requires zero conversion.  Then I tried bumping up to 24/48.  I didn't notice any difference on most things.  Except drums.  For some reason, 24/48 seemed to capture drum sounds with much more accuracy.  At 16/44.1, I was pulling (what's left of) my hair out trying to get the "whole" drum sound, but at 24/48, it all came together.  Or came together with much higher fidelity.  Of course, THAT sparked the "I should be tracking at 24/88.2!!" thing. Do you think that higher sample rates better capture percussive transients?

Isn't this somewhat 'apples to oranges', since you're comparing different bit depths as well as different sample rates? The potential benefits of digital production at 24 vs. 16 bit aren't in dispute -- maybe that explains what you heard.

[Dave Scoven]

A new subtopic here, por favor --

If you are mixing at 24/48, then how far down can you pull a fader before you start to loose signal quality?

What I've been doing for a long time is anticipating where things will be in the mix before tracking, and recording primary instruments so that they find their place in the mix with the track fader as close to 0db as possible.  I started doing this because I could hear signal degradation as I pulled down the channel fader.

I've been doing a bit of reading on the Web and have come across a couple of mentions regarding this phenomenon.  Essentially these articles say that zero loss is attained by an inverse relationship between fader position and word length used during mixdown -- that is, the more you pull the fader down, the higher the bit "depth" needs to be.  So if you want to retain the signal integrity at -30db, you need to be mixing at X bit depth ...

For me, if all of my faders are close to the 0db position, I guess the point is moot.  But for folks who need to pull that acoustic guitar track down to -20db, this could be an issue - no?

[Tomas Danko]

It's down to the mixing engine/summing bus. If you're on ProTools or using something else that uses 32 bit floating point instead it's a rather moot point because you can pull down them faders until the cows come home and still retain the initial resolution of the audio data.

It's the other, upper, end of the meter one should be aware of.

[bruno putzeys]

If the internal buses are fixed point, the adverse effects may still be limited to noise if correct dithering is used to round intermediate results. To check that, simply attenuate the signal 90dB and boost it by 90dB. Then do a null test between the original and the new signal (invert one and mix). If the processing is floating point, the result will be quiet. If the processing is fixed point with correct dithering, the result will be steady white noise without a hint of the original audio. If the processing is fixed point without correct dithering, you'll hear something that's clearly correlated with the audio, especially during quiet passages.

Another test that's always worth doing is attenuating and subsequently boosting a signal by only 0.1dB and then nulling. Boost the nulled signal until it becomes audible and listen. You will always hear something, the question being what and at what signal level. Good fixed point systems will only produce steady noise. Floating point systems will always produce something more correlated with the audio, but at a level that scales up and down with the signal, so it's never likely to be audible.

[snaggle]

if you mix to an external device wouldn't the better capture actually translate?

[bruno putzeys]

snaggle wrote on Sun, 04 February 2007 19:31

if you mix to an external device wouldn't the better capture actually translate?

Nope. The finished product will be indistinguishable from one entirely made at the lowest sampling rate in the chain. That's the big difference with longer word lengths, where producing at (much) longer word lengths than the finished product is a must.

[mikepecchio]

I just remembered an experiment I did a while back:

I converted all the files in a 30+ track nuendo V1.5 session from 44.1KHz to 88.2kHz, then digitally mixed to an 88.2KHz file. I converted the mixdown to 44.1 and compared it to a version that was mixed directly to 44.1 from the original 44.1 session.  They do not null.  It sounded noticably different.  primarily, the upsampled/mix/downsampled version had better depth and stereo imaging. I was surprised at the results as I am somewhat "allergic" to SRC.  How could all that extra math sound BETTER?  Could this be attributed to the way nuendo does its summing algorithm?  I ended up mixing a whole album that way.

mike p

[bruno putzeys]

The best candidate to explain this would be the EQ's. It would appear that they are implemented somewhat na

[Dave Scoven]

Bruno, I can see that recording at 44.1 then bumping it all up to 88.2 does nothing but waste space.  I can also see that 88.2 processing of a signal recorded at 44.1 doesn't do anything to improve the fidelity of the 44.1 signal.

But I wonder if mixing at 88.2 does give the effects a better sound.  Not that the original signals sound better, but the verb introduced during mixing sounds better at 88.2 than it would if you mixed at 44.1?  Does that make sense?  Of course then you loose it again when you go back to 44.1, so the thrill is short-lived ... Unless something is preserved when you go back to 44.1...

[bruno putzeys]

The answer is right above your question
(with some further elaboration in my first answer in this thread). I couldn't quickly track down a readable IIR filter tutorial, but you can try your luck on google with key words like "IIR filter", "bilinear transform" and "warping".

[ScotcH]

Dave Scoven wrote on Wed, 07 February 2007 14:09

Bruno, I can see that recording at 44.1 then bumping it all up to 88.2 does nothing but waste space.  I can also see that 88.2 processing of a signal recorded at 44.1 doesn't do anything to improve the fidelity of the 44.1 signal. But I wonder if mixing at 88.2 does give the effects a better sound.  Not that the original signals sound better, but the verb introduced during mixing sounds better at 88.2 than it would if you mixed at 44.1?  Does that make sense?  Of course then you loose it again when you go back to 44.1, so the thrill is short-lived ... Unless something is preserved when you go back to 44.1...

I asked a similar question in the WW forum, and the concesus has been that mixing into a higher sample rate is useless if ITB, but usefull when mixing on a console, then back into the PC for 2 track.  Here's how I see the possibilites, and the conclusions ... please let me know if I have it right (assume all ITB):

1. track at 44.1/24 - mixdown to 44.1/24 --> mastering (back to 44.1/16)

2. track at 44.1/24 - mixdown into 88.2/24 --> mastering (which brings it back to 44.1/16
- It would seem that this does nothing over 1, since the upsampling to 88.2 is point less

3. track at 44.1/24 - upsample to 88.2/24 - mixdown into 88.2/24 --> mastering (which brings it back to 44.1/16
- This is what you're asking about above.  It seems that you gain the higher resolution for plug-ins during mixdown
- Resource intensive (ie space, CPU power, etc.)
- sounds best while mixing

4. track at 88.2/24 - mixdown into 88.2/24 --> mastering (back to 44.1/16)
- This seems like the best chain, but the most resource intensive (ie space, CPU power, etc.)
- gives the best plug-in quality
- sounds best while mixing (if your ears can tell ...)

It would appear that 1 is OK but not optimal, 2, offers no benefit over 1, 3 is better to benefit from the better resolution, and if you do 3, might as well do 4 since it's the same resource requirements.

I believe 3 & 4 would sound better than 1 and 2 since you preserve the plugin resultion.  Remember that these act on the individual tracks, and are then mixed down.  The final mix is then down sampled.  That's the differnce it seems to me ... if you processed each track individually, then downsampled and mixed, I could see how the result would be identical to 1 (and 2).  The biggest issue for me right now is resources (CPU bogs after 25 tracks or so) so I'll be sticking with 1 for now

[bruno putzeys]

It's a bit quick to say all plugins will work better at higher sampling rates.
Dynamics will work better.
EQ's will work better when working higher frequencies but worse when working low frequencies (with sufficient precision arithmetic this should not necessarily be a big problem).
Reverb etc will simply eat more resources.

Presuming the EQ has no word length issues, we could say that 3 is better than 1. If the EQ's take account of response aliasing in their design (e.g. Oxford), 3 and 1 will sound the same. Which demonstrates that improved filter algorithms make more sense than upsampling the whole show.
4 may sound better than 3 during tracking and mixdown but will lose all advantage over 3 when converting back to 44.1k. However, 4 does not use more resources than 3.

[Sin x/x]

Bruno Putzeys wrote on Thu, 08 February 2007 03:21

Dynamics will work better.

Haven't designers ever heard of locally up-sampling and down-sampling?
This will eliminate the need for higher sampling rates.

[bruno putzeys]

Some have, some haven't. Dynamics with local upsampling in the side chain do not benefit from further upsampling. The point was already addressed in the second post in this thread.

[Sin x/x]

Bruno Putzeys wrote on Thu, 08 February 2007 08:53

Some have, some haven't. Dynamics with local upsampling in the side chain do not benefit from further upsampling. The point was already addressed in the second post in this thread.

Ok didn't read that.

But it does mean, that with good designed plug inns higher sample rates are meaningless. So why do people put up with that crap.

Unless people evolved to hear above 20kHz. And there's no scientific proof of that.

[bruno putzeys]

Sin x/x wrote on Fri, 09 February 2007 06:20

But it does mean, that with good designed plug inns higher sample rates are meaningless.

If the plug-ins are done right, higher sampling rates than the release rate are pointless.

Sin x/x wrote on Fri, 09 February 2007 06:20

So why do people put up with that crap.

The same reason why audiophiles buy expensive RCA cables, mains conditioners, DACs that slave to SPDIF and fancy transports. If you start with a system that is "broken" (unbalanced connections, slaving AD/DA to an embedded clock), there's a lot you can do to change the result. That gives a sense of control. Suppose you install a system with balanced I/O and you swap cables to show that there is no longer a difference. Aww man that's bo-o-oring. Took the magic right away. Even if a system that is insensitive to this parameter actually sounds better, they'll prefer one they can tinker with, because it puts them in a "privileged position". This is why automatic transmissions have never caught on in Europe. People here all believe they can shift gear better than an automatic transmission can. Yeah Schumacher perhaps.

Sin x/x wrote on Fri, 09 February 2007 06:20

Unless people evolved to hear above 20kHz. And there's no scientific proof of that.

Don't go overboard... theory of evolution itself does not preclude hearing above 20kHz. It's just that it happenend to work out that way for homo sapiens.

There is scientific proof that perception doesn't stop dead at 20.000kHz. A simple ABX between half-band and nonaliasing upsampling filters in a 44.1kHz system shows that at least the range between 20k and 24.1k affects perception provably. You can try it for yourself: make a nice loooong filter cutting off very sharply at 20k and insert it in a 44.1kHz chain. This leaves everything below 20k intact while removing the muck left by half-band filters (that don't even fulfil the Nyquist criterium). Or use 48kHz and make a shorter one flat to 20k and first zero at 21.6k.
If we had standardised to 48kHz and not used a halfband filter for the first upsampling / last decimation stage we wouldn't have had to put up with the rate race. Because, to get back to your original point, there is of course enough scientific evidence that the buck does stop somewhere very close to 20kHz, certainly not one or two octaves higher.

[Tomas Danko]

Hi Bruno,

For what it's worth I just wanted to take the time and congratulate you for your natural way to simplify and explain about misbeliefs when it comes to audio. I see a little Dan Lavry in you, and I like that very much. You seem to have a very sober view on things, based on empirical facts, but you've maintained an open mind nevertheless.

Furthermore, you have managed to explain all these things without stirring up emotions that turn into pointless flaming. That makes you a great moderator.

You have my respect. Have a Rodenbach Grand Cru on me this friday, you sure deserve it!

Sincerely,

Tomas Danko

[bruno putzeys]

Thanks

[Yannick Willox]

I object.

Gearshifting at the very least has a Freudian implication. Besides that, there are big energy losses in non-robotised automatic gearboxes. A car with a manual gearbox with a good driver will consume considerable less than one with a 'standard' automatic gearbox. And accelerate quicker.

But I agree completely on the higher samplerate issue.

[EP]

Bruno Putzeys wrote on Thu, 08 February 2007 01:21

4 may sound better than 3 during tracking and mixdown but will lose all advantage over 3 when converting back to 44.1k. However, 4 does not use more resources than 3.

I have enjoyed reading this through. Although slightly OT, to me it bears bringing up that having your original tracking/mixing sessions done #4 style (all 88.2) has an additional benefit:

44.1 is a marginal spec that thrives because of the CD (and mp3) that should both be well and dead soon. Wouldn't it be nice to start putting out download versions of your music at 88.2/24? Already this is playable by anyone with a computer. If not now, soon this may be a real option. If tracked/mixed at 44.1 you can't offer the 'improved' (original   ) version.....

Erik

[bruno putzeys]

I would already be willing to pay a premium to download a non-squashed version - mastered and edited but with the full dynamics (and peaks) intact.

[DigiEm]

EP wrote on Sat, 10 February 2007 12:55

44.1 is a marginal spec that thrives because of the CD (and mp3) that should both be well and dead soon.

I think MP3s have their place.
But Red Book Audio-CD is such a funny little standard that should be dead long time ago.
If only CDs had their data recorded in CD-ROM format or would have a second layer
of data in CD-ROM format which CD-players could rip in seconds and play back from memory.
It would both improve the sound and make audiophile quality players cheaper to manufacture.

By the way, what about flash based MP3/WAV players?
Aren’t they potentially the best playback devices?
I mean aren’t they the least jittery playback source?
Also flash based players can be powered by a small rechargeable battery for many hours.
I understand that rechargeable batteries provide the best and cleanest power.

Do these qualities like flash memory and rechargeable battery make it easier and cheaper to
design audiophile quality converters for flash memory based players
as opposed to CD-players and stand-alone converters?

[Dave Scoven]

If the standard shifted from 44.1 to 88.2, wouldn't that render 90% of home studios' computers totally worthless?  I mean, I don't know what it would take to run 16 or 24 tracks of high quality plugins (especially plugins like Altiverb) at 88.2, but ...  LOL -- Bruno??

[mikepecchio]

Bruno Putzeys wrote on Fri, 09 February 2007 03:05

Even if a system that is insensitive to this parameter actually sounds better, they'll prefer one they can tinker with, because it puts them in a "privileged position".

I agree totally.  same goes for power amps with significant output-Z and low dampng factor. for some reason loads of people WANT to hear their speaker cables.

[EP]

Dave Scoven wrote on Sun, 11 February 2007 08:16

If the standard shifted from 44.1 to 88.2, wouldn't that render 90% of home studios' computers totally worthless?

To my thinking this scenario has always existed. In days past the home studio had to make do with a tascam tape deck, where the pro's used a Studer....

I would rather see the standard at the higher s.r., and those that can't produce at that level SRC up to meet the standard rather than the recordings that can take advantage of the higher S.R SRC down to the lowest common level......

Erik

[Garrett H]

In mastering, we often get mixes in 24/44 or 24/48.  Those of us using digital gear (outboard digital, not necessarily plug-ins) such as Weiss and TC Electronics System 6000 and Z-sys, et al) have found that upsampling the source to 88 or 96, doing all mastering work there, and doing one SRC and one Dither down to 16/44 at the end ...(is this a long sentence or what?) usually sounds better in terms of perceived front to back depth of field, stereo imaging, and the ability to place instruments in a specific place in the playback field.  How much better this sounds always depends on the type of source (things recorded with room mics and ambient mics) usually gain more from upsampling.  Conversely, sample-based productions gain less.  

Jumping around in the mix stage is an intriguing question.  And please forgive my nastiness in the following suggestion, but why don't some of you do some tests with material you work with and decide for yourself.  By all means, please report back to the group - we grow as a community that way.  But sometimes you need to try this stuff yourself.  Some of the people on web boards don't ever EVER do any recording.  They're just professional web posters.  ( And I'm definately not point a finger at anyone on this thread).  But you got to fall on your own to see how to keep your balance in the future.

One last comment on 16/44 etc... I've read that this standard is way beyond sufficient, but our A/D and D/A converters and their filters are not good enough to take advantage.  Using 24/96 allows sloppy designs to sound better.  Does anyone know any more about such claims?  I'm not up on converter design enough to argue one way or another at this point.    Oh, heck, who wants DSD?

Best,
Garrett H.

[bruno putzeys]

Garrett H wrote on Mon, 19 February 2007 05:43

(...)that upsampling the source to 88 or 96, doing all mastering work there, and doing one SRC and one Dither down to 16/44 at the end ...(is this a long sentence or what?) usually sounds better in terms of perceived front to back depth of field, stereo imaging, and the ability to place instruments in a specific place in the playback field.  How much better this sounds always depends on the type of source (things recorded with room mics and ambient mics) usually gain more from upsampling.  Conversely, sample-based productions gain less.

More precisely and more generally: A/B miked stereo gains significantly, panpotted or intensity stereo improves hardly, if at all.
We're once again back to the old half-band filter problem. Half-band filters are the natural choice for an upsampling/decimation chain except for the 1*fs -> 2*fs (or vice versa) stage. Nyquist requires no energy to be present at or above fs/2 at sampling, and requires all energy at or above fs/2 to be removed at reconstruction. Under those circumstances, the sampled system is indistinguishable from a low-pass filter. However, for economical reasons that lowest stage is almost invariably half band too, with the pass band to 0.4535*fs (yes: 20.000kHz at fs=44.1kHz)and stop band starting from 0.5565*fs. The attenuation at fs/2 is exactly 6.02dB. This means that between 20kHz and 24.1kHz there is a region where imaging and aliasing reign. To do the test, get a square wave generator and a scope. Feed the square wave into the AD converter and see what comes out of the D/A. Whenever one of the square wave's harmonics is inside the 0.45fs to 0.55fs band you'll see the pre/post ringing wobble. Suppose that square wave is a sound source somewhere in the stereo image. In the case of intensity stereo, the square wave looks the same on both channels. The aliased components will come from the same spot in the stereo image as the main source. In the case of time delay stereo, the alias products will be different on both channels, and they will sound as coming from elsewhere in the stereo image. Clearly the latter will be much easier to hear.
Step 2: insert a x2 SRC followed by a /2 SRC. The wobbling will be nearly imperceptible on the scope. The SRC uses 0.45/0.55 half-band filters too, but you've passed the signal through two of them, which adds a further 12dB of attenuation at fs/2.
Do this experiment with music too: start with a 44.1k original, upsample to 88.2 and back to 44.1 (without any further processing) and listen for the difference. You might think you're only converting sample rates, but it's the filtering this entails that makes the difference, not the sampling rates.
The intelligent solution is not to do SRC but to insert a low-pass filter that does exactly what we want: remove the band above 0.45fs.

[bruno putzeys]

Discussion relating to DAW hardware specs for high sample rate work hsd been moved to a separate thread called "DAW hardware recording bottlenecks" in the "Recording hardware products" forum.