Pro Tools Sound Quality???

[Rail Jon Rogut]

Mike

You don't think it's time to replace the old 24 bit TDM buss with a newer 64 bit TDM buss with your next major hardware upgrade? This would certainly take care of the problem.

As to the blind test -- I doubt that any results you garner from a blind test performed under varying conditions by varying users of whom we know nothing about their qualifications, will give any valid results. You may want to do what they did with the watermarking technology -- Warner Bros. held A-B-X listening tests at different areas of the US with professionals in the audio field including members of MPGA (Ed Cherney, Massenberg, etc) and mastering engineers (Doug Sax, Bob Ludwig, bernie Grundman) -- and based upon those results they could make a qualified assessment.

I also think another valid test would be to perform these tests in a professional recording studio where you can A-B between a full mix where the outputs of Pro Tools are taken individually out of Pro Tools into faders on an SSL 9k and summed using the 9k's summing buss vs. the stereo mix summed inside Pro Tools.

Best regards.

Rail

------------------
Recording Engineer

[Rail Jon Rogut]

Hi Monte

I agree that the dither in the current TDM buss will probably solve most of the ills - but would consume CPU time in processing... so I thought the logical step would be to make the wider TDM buss, which wouldn't require any additional processing overhead (well perhaps some to handle the larger bit width numbers) - thus allowing Digi to maintain their current track count.

I also agree that the floating point math in the 001 may indeed sound superior, but as George Massenburg told me, you certainly can get acceptible sound quality using fixed point if it's done correctly. He felt the main culprit was the truncation -- without that the quality should improve.

Best regards.

Rail

------------------
Recording Engineer

[Monte McGuire]

Mr. Rockwell,

Thanks for taking this subject seriously. I really look forward to such a test, but I'd like to make a few suggestions as to how to approach the whole thing.

The problem I have with the sound of the PT mixer happens when I have a large number of inputs summed to mix. It's not that the PT code is misbehaving in some sort of bug like way with large mixes, I think it's simply a case that emphasises the problems with the current mixer. So, any test files made with the two mixers should be made from a typical large scale mix, and not just a few tracks run through the mixer.

The tracks themselves should be sourced through a reasonably clean chain too. If a number of plug ins that round or truncate back to 24 bits are used as inserts on each channel of the test mix, the importance of the final truncation at the end of the stage will be minimized and any differences won't be as audible. We'd simply be comparing twice damaged audio to thrice damaged audio and yes, the differences won't be that significant. So, restrict plug in usage to those plug ins that do 48 -> dither to 24 processing only.

I'd also like to suggest that you use artificial reverb in the mixes. Reverb gets altered by truncation pretty readily, and it's therefore a good test of processing quality for both the aux send mixers and the main stereo mix. Truncation at either end will shrink and harden the sound of the reverb.

Finally, I would prefer it if you could post two unidentified mixers instead of the mixes themselves. You know, mixer A and mixer B and let us decide whether we can hear differences and decide which is which. I guess it would be readily apparent which mixer is which if you looked at allocator, but maybe there's some way to make the non dithered mixer out of the dithered mixer by patching out the dither code. In this way, the two mixers would use up the same DSP resources and be indistinguishable except for their processing quality.


Again, thanks a lot for taking this seriously!! I look forward to the tests...

Monte McGuire
[email protected]

[Monte McGuire]

Rail,

I don't think a super wide TDM bus is needed if every processing step dithers the accumulator before going to the TDM bus. In that case, the only difference would be the small resolution loss from adding dither at the 24 bit level, and that is a very small increase in noise. Remember, true uncorrelated noise is relatively benign, but the cross modulation distortion products caused by truncation are not.

To put some numbers on it, assume you have three inserts on each track of a 32 channel mix. The added noise of the dither between each plug in amounts to a noise floor of around -140dBFS per track. Summing 32 of these tracks at unity gain yields a noise floor of -125dBFS.

Currently, that's about as quiet as a single channel of conversion, so I don't think it's really going to have a huge impact. Plus, if the mixer worked properly and you could use the faders to attenuate individual tracks, then the resulting noise floor could be a good bit lower - there is no need to run all the faders at unity gain and the noise floor could drop quite a bit.

However, if dither is not going to be used, then yes, a wide TDM bus would help a lot. It seems to me that this is a great argument for dither in that we could get the same sound quality in a 24 bit chain as we could from a true 48 bit path, plus a little extra noise.

One final comment about a wide TDM bus. Native floating point based systems are usually undithered between processing steps and they have an internal chain that is essentially the same width as their 24+8 bit accumulator. Given that the new Digi 001 system uses only floating point processing and has session file and plug in compatibility with PT TDM, it would be a great test to run the same mix on an 001 system as well as the current TDM mixer and a new dithered TDM mixer.

Since floating point processing will clean up the plug ins as well as the mixer, I wouldn't be surprised if a mix sounds best on the 001.
I know I have been pleasantly surprised by the quality of work some of my friends are doing on low end floating point systems and frankly, if the editing and latency on those systems were better, I'd have jumped ship!


Monte McGuire
[email protected]

A 64 track session would be an ideal test example!
Jules



[This message has been edited by Julian Standen (edited 10-03-99).]

[Digi Engineering]

Thanks for all of your feedback. We'll take it into consideration when we create the test files. A couple of comments:

Monte: In your post you state that 32-bit floating point is 24+8 bits worth of precision. This is not actually the case. With 32-bit floating point there are actually 24-bits worth of mantissa. The other 8-bits are for exponent. As far as I know all of the 32-bit host systems out there normalize their values to between -1 and 1. Anything greater than 1 is considered clipping. This is actually less resolution than the 48-bit fixed point that is used in the PT 24-bit optimized mixer.

Disco: It's unfortunate that you think that we would manipulate the results. We will post the honest results and descriptions of the techniques used to create the sound files at the end. I hope that you will trust us to be honest. This test is really to help us decide the best way to make PT sound better not to be able to make some marketing claims.

Rail:
Since we can probably never make our mixer sound like the SSL (analog vs. digital) it doesn't seem like a very useful test. If there is a digital console that you think sounds a lot better than PT that would be useful info. Extending the width of TDM would actually increase the processing overhead substantially. 56k DSP's have a 24-bit external data bus. It would double the overhead to write 48-bits to a wider TDM bus. Again I think that it's easy to get caught up in the more is better syndrom. 64-bit fixed point would have a 384dB dynamic range which would basically be mostly a waste of bandwidth. Let's see how the tests come out before we jump to that conclusion.
Thanks,
Mike Rockwell

[J Harry]

Mike,

It would be great if you used an SSL Axiom

JH

[Disco_Doctor]

It's unfortunate that you think that we would manipulate the results.

Mike, why do twist around everything I say? I said "in the interest of making sure that the test is not manipulated by either Digidesign OR ITS DETRACTORS". That way, no one at Digidesign can ever misrepresent the results, and no one who disagrees with or challenges the results can accuse Digidesign of manipulating the results.

I did not say, nor do I actually believe, that you or any other particular person would use the test for marketing purposes. But just because YOU are an honest guy with a good intentions, that doesn't mean everyone in Avid's boardroom and marketing department is. We're talking about a huge corporation here Mike - of which you are not the CEO or CFO - and you can't, with a straight face, speak to the motives of everyone who works there. My lack of unequivocal faith in Avid Inc. is not "unfortunate" - it's pragmatic.

If you want this test to have long term integrity, the "answers" should be given to a third party. I stand by what I said.

I doubt that any results you garner from a blind test performed under varying conditions by varying users of whom we know nothing about their qualifications, will give any valid results.

Maybe so. But I also think that with a public test like this, the listener results may average out and provide some useful information. There's plenty of qualified ears out there in Digi's customer base. Your suggestion is good though - Digi should also subject itself to the opinions of MPGA members and top mastering engineers as well as its broad audience of amateur and professional customers.

I'm sure there will be plenty of situations where producers and engineers will get together with their peers to do this testing in professional studios. That will lend some additional validity to this public test.

One other suggestion Mike. Set up a temporary forum on the DUC specifically for the test results. Post instructions that outline exactly how users should perform the testing and a questionaire that lists specific questions for the respondents to answer. That will make it a lot easier for Digidesign to extract useful data out of all the feedback.

[Rail Jon Rogut]

My thought was that every time the signal has to enter a TDM node, it also has to be dithered down to 24 bits, and at that stage it's done by the CPU - dithering on returning out of the plug-in or TDM mix node would be done by a DSP chip... of course I have no real knowledge of this and it would be up to Mike to let us know exactly where and how the dithering was being performed. You may be correct... I just don't know (it would be nice to have a block diagram one day).

Regards.

Rail

[Digi Engineering]

It is true that you need to have head room to mix together phase synchronous audio. This applies to analog as well as to digital. Analog consoles let you go to around +28dBU before you clip the stereo bus. The 24-bit optimized mixer has around 30 dB of headroom above zero before it clips internally. This was not the case in the past with the 16-bit optimized mixer which would clip pretty easily and you had to bring down the source faders. With the 24-bit optimized mixer you can bring down your master fader if you are clipping the output, just like you would do on an analog console.
-Mike Rockwell