Mixing Engine exceels on T2

[M'Snah]

I probably miss something. GYang doesn't have to prove to me. He notices a clear difference between different bitdepth mixings. I have no problem believing that. Why should he prove it? If you don't believe it, just keep on mixing at 32bit. Unless you need this 'evidence' as a persuasion to buy Tracktion 2?

[.jon]

Well.

I couldn't judge a sound by it's waveform, but I'm going to try to hear the difference myself.

Hi-end audio things are always subjective, and GYaang just expressed his personal experience... I'm with HansM, no need for proofing here.

[stephengardner]

I am in the final mix stage of a 24 track plus project.. 11 songs using 2 UAD cards Tannoy monitors on T1.. I will print version 1 32 bit and then load T2 and print version 2 64bit.. It is being mastered by Randy Leroy at the Final Stage Nashville I will report his comments on the difference.

[semiquaver]

yay stephen!

[Barbarossa]

stephengardner,
to get a valid result, you have to use 32 and 64 bit on the same audio engine, since we don't know about changes between T1 and T2 (i.e. automation updating).

So only use T2 and make sure any automation and other time-invariant effects are switched off (i.e. volume automation which is not sample-accurate in T, chorus-effects, non-impulse reverbs, VSTis,...).

[willowhaus]

I'd be interested as well in what the test shows. Hearing is the bottom line, but it is also very subjective. Even blind listening tests are suject to variations because of this.

Funny - I was just reading a slightly old thread here yesterday, with people ripping on a very well-known audio engineer who preferred Nuendo because "it sounds better" and offering up all sorts of theories as to why he was wrong...

[IIRs]

So only use T2 and make sure any automation and other time-invariant effects are switched off (i.e. volume automation which is not sample-accurate in T, chorus-effects, non-impulse reverbs, VSTis,...).

This is only relevant for the phase invert / sum test. Why would he do that at a mastering studio?

[GYang]

Guys,

I am not against proofs, but to be clear, no better proof is than listening to your own projects (again, if you have appropriate setup).
Technically it is possible to make two 3 min. masters in 24-bit format and put on Internet for comparison, but not with mentioned song where work is in progress and it is simply not my own material.

The point is different.
I used Samplitude and Nuendo as reference sequencers for couple of years and know very well Pro Tools and its sound especially.
Samplitude was slightly better sounding than others (although there are many who would say that there is no sonic differences between DAWs in terms of purely mixing if everything else is same).
I don't know what was done to T2 mixing algorithm, but T1 project opened there (in 64 bit) immediately caught my attention. It was not my 'auto suggestion' or 'wish to hear magic whispers'. It just happenned that slightly unclear passages in complex mix all of sudden improved, became transparent. I checked if something was changed, but none.
Next days that repeated, so it is not question of good mood or specific conditions.
Listening on second set of monitors (mid priced Tannoys) still shown improvements, but much less, so I concluded that many users with weaker monitoring section would simply not hear it.
More interesting is that mixing in 96 or 192 kHz sample rate gives less benefits with 32 bit than 48 khz with 64-bit mixing. Unfortunatelly today's PCs do not give us enough power to record and mix in 24/96 or 192 with large tracks number, but I could easily imagine that next generation of 64 bit Windows will allow us to reach this target and come near to perfect audio on (advanced) home studio level. So, for the time being T2 offering is awesome in that respect and probably some newcomers from other expensive platforms (as me) to T2 are quite possible reality.

So, my suggestion is that all interested should try and give their honest opinion.
If more than 10% would hear no difference, I'll undertake to record one experimental song and put it on Internet for public judgement.

BTW one can still find many opinions on specialized forums that SACD is not sounding better than CD (and in pure mathematics it could be proven - please, don't ask me to give you reference proofs on this issue, too).
But in some kind of comparison I could say that difference in T2 64 vs T2 32 is something like CD vs SACD. On some material this difference is drastic.

I would dare to say that for the time being T2/64 is most probably the best sounding audio mixing software regardless price and market segment (and that is quite achievement).

yang

[Barbarossa]

Yang,
i don't trust in what you claim to hear. Even less if the monitors are Tannoys.
I only trust in what i can hear myself.
And if i can't check it, i'm free enough to use my mathematical and technical knowledge to check, if it makes sense, what is claimed.

And i'm sorry: it doesn't make much sense to me, if you claim to hear a difference in sound quality, if 20 tracks are simply mixed together in 64bit float and afterwards rounded to 24 bit, compared to the 32bit-float mix.
The 32bit-float format is precisely enough, to add 20 24bit-fixed tracks without any rounding errors.

Can you explain to me where there should rounding errors appear or where the 64bit calculation can give a better result?
1+1=2
1.0+1.0=2.0
There is no difference.

That's not a matter of believe, that's a matter of empirical, mathematical facts.

64bit can indeed have an influence, if there are tiny enough amplitudes, becoming lost in the 32bit format (think of reverb-processing) that the addition of that rounding errors could become audible.
But where could they appear in your simple mixing situation?

I don't want to discuss the example with SACD vs CD. It should be clear what the difference between SamplingRates and BitDepths is and i was not claiming that there can be no audible differences. Two pair of shoes.

You claim to hear a difference. I doubt it.
So convince me either with an example or with a mathematical argumentation, that it could be at least theoretically possible.
That's all i want.
But to claim "i can hear it" and "i know so many other applications sounding worse" is a bit too less for convincing a skeptist.

But nevertheless i will test it one day on my own anyway.

[semiquaver]

The 32bit-float format is precisely enough, to add 20 24bit-fixed tracks without any rounding errors.

Barbaeossa, I'm sorry but this is simply nonsense. There is rounding error in each and every gain change at any bit depth in floating point math. it is not a matter of 1.0 + 1.0 = 2.0 unless you use no processing and mix all tracks at unity gain.

In fp every gain stage adds again the default rounding error - and errors in earlier stages can then be amplified in later stages. - thus if you cut and then boost a signal in 32 bit by 20db you will get rounding noise at around 20 db above the theoretical noise floor (if I'm not mistaken that would be around -84 db - not negligable). Using compressors and eq plugs (not to mention vstis which can compound these errors via many small notes) it could be perfectly possible to introduce a level of noise that could persist into your 16-bit master.

You should also consider the considerable error that filters (ie eq) can introduce.

[.jon]

Can you explain to me where there should rounding errors appear or where the 64bit calculation can give a better result?
1+1=2
1.0+1.0=2.0
There is no difference.

1+1=2 as opposed to 1.4+1.3=2.7

Probably not very much related to this discussion about the said benefits of 32 versus 64 bit mixing, but that's the rounding error _you_ are talking about

[MT]

I think there's one of these "fuzzy math" threads on every f-ing audio forum out there. So tiring.

MT

[stephengardner]

Yang..I am glad you are doing these camparison of T to Protools, Nuendo and posting info here..

[GYang]

And i'm sorry: it doesn't make much sense to me, if you claim to hear a difference in sound quality, if 20 tracks are simply mixed together in 64bit float and afterwards rounded to 24 bit, compared to the 32bit-float mix.
The 32bit-float format is precisely enough, to add 20 24bit-fixed tracks without any rounding errors.

Can you explain to me where there should rounding errors appear or where the 64bit calculation can give a better result?
1+1=2
1.0+1.0=2.0
There is no difference.

Barbarossa,

it would be easier to invite you to convince yourself in my studio than to learn all math principles and prove it.

Well, back to your question.

Let me use original words of today's one of recognized top five person in digital audio technology who participated in creation of early Apogee converters, as well as, who produces one of arguably best converters (Lavry Gold) in the world. For more details you can participate in his forums, I would be happy to finish discussion here:

danlavry wrote on Mon, 18 April 2005 22:06

'It is fine to use a lot of bits for digital signal processing (the "mix bus"), such as for adding a lot of channels with gain, attenuation, filtering...

When does one begin to need a 64-bit bus for mixing 32-bit signals (not for filtering, I mean multiplying and summing)?

...' But say you want to "set" your single channel of 24 bits data "at the middle" of a 64 bit bus. So 64-24=40. That allows you to have 20 bits "above" and 20 bits "bellow". In other words you have 00000...XXXXXXX...00000, where the X's are music data. That by itself allows you to digitally amplify and attenuate by about 120dB (shift by 20 bits). Now lets add 32 channels of 24 bits at the same XXX location, and you really have 5 more data bits... But if you applied gain and attenuation prior to accumulation, the margins are reduced.
Of course, arbitrary multiplication constant will produce LSB's all the way down to bit 64 and the ability to truncate after addition (instead of adding of truncated words) yields more precision at the final LSB level.
64 bits may be a bit much for limited channels, but having a nice margin allows the operators an easier time at avoiding clipping and truncation.
Also, I do not know if the software guys need to compute power. Do they need such computation for sat a compressor algorithm? I do not know. But if they do, than the X^2 operator will double the number of required bits.
In any case, 64 bits is getting to be "sort of a standard" with some of the major players. It may be a bit of an overkill, but it sure beats 32 bits which was too limiting...

Regards
Dan Lavry
www.lavryengineering.com

Actually there are many more elaborate maths opinions on 64 bits (I hate maths really, so don't wish to pollute this environment more).

Peace Barbarossa and enjoy life more.

yang

[Barbarossa]

The 32bit-float format is precisely enough, to add 20 24bit-fixed tracks without any rounding errors.

Barbaeossa, I'm sorry but this is simply nonsense. There is rounding error in each and every gain change at any bit depth in floating point math. it is not a matter of 1.0 + 1.0 = 2.0 unless you use no processing and mix all tracks at unity gain.

In fp every gain stage adds again the default rounding error - and errors in earlier stages can then be amplified in later stages. - thus if you cut and then boost a signal in 32 bit by 20db you will get rounding noise at around 20 db above the theoretical noise floor (if I'm not mistaken that would be around -84 db - not negligable).

But this is not correct for the FP-format.
There appear only rounding errors, if the result can not be represented.
If you reduce the level, say 50% or 6dB, then simply the exponent is reduced, while the whole mantissa value keeps it's value.
Reducing the level 10 or 100 times, simply results in a different exponent without losing any information in the mantissa.
Increasing the level, then changes the exponent again.

Using compressors and eq plugs (not to mention vstis which can compound these errors via many small notes) it could be perfectly possible to introduce a level of noise that could persist into your 16-bit master.

True, but we are talking about a claimed clearly hearable difference from the 24bit master-output, when switching between 32 and 64bit.

You should also consider the considerable error that filters (ie eq) can introduce.

Yes, i said so, but really heavy processing is necessary for that to reach an audible level, not one EQ stage per track.

@GYang:
instead of a neverending theoretical discussion, you could simply make the test in two minutes, by rendering and comparing the results.