Print

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

[Pingu]

Debate..............

That might all be good in the implementation of the algorithm and
in theory.

My experience with SRC and audio has led me to believe that you cant accept statements such as "use an even multiple of your destination format as the SRC will be sonically superior to that of an uneven multiple", at face value.

[Matt_G]

danlavry wrote on Fri, 07 July 2006 03:00

Hi Matt I would certainly separate the number of bits from the sample rate. First about bits: In the good old days, we were struggling to have clean 16 bits. The analog noise floor was so high, that it served as dither. At very high noise, referenced to the quantization level, the dither action was fine, but the noise penalty (reduced dynamics)was too high. Next we started having the noise approaching the 16 bit quantization, eventually having the noise lower then the 16 bit quantization. At that point, it became necessary to do something about the unwanted "art effects" (distortions and noise modulation) of low level signals. By now, many AD's are better then 16 bits. In fact, when an AD is a "true 18 bits" or more, and the end result is say 16 bits, it is best to dither. There is no reason to record directly to 16 bits, when you can record to 24 bits (the last few bits are of no value, they are noisy) and then dither to 16 bits. Of course, a noise shaped dither is better then the older varieties, because the added noise resides in frequencies where the ear sensitivity is lower. As a rule, keep the word length as wide as you can (say 24 bits), and dither at the last step of the processes (with a noise shaping type if you can). Regarding the sample rate: take 44.1KHz vs. 88.2KHz. This days, virtually all converters "do their job internally" at much higher rates then 44.1 or 88.2... Therefore, having 44.1KHz requires some down-sampling to take place. So the first question that come to mind is: which is a better downsampler? The one in the AD? An external SRC? This is important because the last stage in the down sampling process is the one performing 88.2 to 44.1KHz. That last stage is the most difficult and demanding one (in terms of down sampling). So the answer depends on specific implementations of the SRC...

Thanks for your reply Dan, this makes total sense. I guess there is another advantage in capturing at 44.1kHz through the converter & that is to do with clipping the input. If you have a clipped 88.2kHz file that is at or just under 0.0dbfs & you do a downsample using any high quality SRC, you end up with intersample peaks that could exceed 0.0dbfs resulting in nasty distorted artifacts. At least when you clip an A/D input set to 44.1kHz you no longer need to worry about another DSP process that will distort your audio further. Would I be correct in stating that?

I've found that any further processing applied to clipped audio will sound extremely bad, so I am wanting to eliminate any unnecessary steps after the clipper to prevent this. It sounds ironic, but there is clean clipping & then there is the nasty crackling type.

From what I have read & understand, the Sterling guys run there analog mastering chain & then capture/clip at 44.1kHz using either the old GML A/D or in Ted's case the Lavry Gold A/D. Both converters seem to clip very cleanly. Ted seems to lower the output to -0.3dbfs to help prevent intersample peaks that can be a problem in cheap quality CD players or D/A converters. The result is clean & loud without obvious distortion. Further analysing the majority of the big name ME's work reveals a ruler flat line with no SRC intersample peaks, which could suggest that this method is widely used now.

Matt

[Jon Hodgson]

Pingu wrote on Sun, 09 July 2006 22:46

Debate.............. That might all be good in the implementation of the algorithm and in theory. My experience with SRC and audio has led me to believe that you cant accept statements such as "use an even multiple of your destination format as the SRC will be sonically superior to that of an uneven multiple", at face value.

How about

"If your SRC sounds worse at en integer multiple of your destination format than at a non integer multiple then the bloke who programmed it made a mistake"??

That you CAN take at face value.

[Pingu]

if you say so

[Jon Hodgson]

Actually I shouldn't have used the term "sounds worse", because that gets into subjective territory.

I should have said something like "has greater distortion from the ideal". This is measurable since what we're talking about here is completely quantifiable. A perfect SRC preserves all components below the new nyquist exactly, loses all components above the new nyquist, and introduces nothing else.

[blue2blue]

Jon Hodgson wrote on Sun, 09 July 2006 13:00

Pingu wrote on Sun, 09 July 2006 16:50 Youve got the jist of it but i still dont buy into the even integer samplerate debate yet. It sounds logical but processing audio does not always follow this path. I can't see that there is any debate, if we're talking about signal accuracy as opposed to possible subjective preference. If your constraint when designing your filter is memory for coefficients, then the 2:1 samplerate conversion will be superior, because it can use half the available coefficients per phase. Any other ratio requires more phases, and therefore fewer coefficients per phase. If your constraint is processing cycles, then the worst case scenario is that the 2:1 conversion is equally good. The truth of the matter however is that given a good quality converter and high quality SRC I think it's unlikely that people could pick out the difference, especially when dealing with real music.

Thanks, Jon!

Since the implication of my question was as to greatest possible waveform replication accuracy for SRC, that answers that question nicely.

And on the practical (though subjective) side, your final paragraph's proviso should give some reassurance to those who need to service multiple conflicting formats, ie, video at 96 kHz and conventional CD audio at 44.1 kHz.

[Jon Hodgson]

blue2blue wrote on Mon, 10 July 2006 16:10

And on the practical (though subjective) side, your final paragraph's proviso should give some reassurance to those who need to service multiple conflicting formats, ie, video at 96 kHz and conventional CD audio at 44.1 kHz.

The most important thing is to get a good SRC. I was quite surprised when I saw this page

http://src.infinitewave.ca/

Now I can't confirm the quality of their testing, but if they've done things correctly then it seems to tell us three things

1) It is possible to achieve extremely high quality (170dB is 28 bits worth)

2) Plugins and programs aimed specifically at mastering tend to have higher quality than your standard DAW SRC (different priority, in a multitrack DAW you are trying to keep processor use down so you can have as many tracks and plugins as possible, in a mastering program you can throw all your resources at just two tracks).

3) Even in the worst case in these tests the artifacts are only just edging up into audibility, so use those CPU draining SRCs when you're mastering, by all means, but don't lose sleep if you don't have one.

[Barry Hufker]

I showed the results of this test to the folk at Audio Ease.  They said the Izotope was operating at 64 bits (as the name plainly shows in the test) and Barbabatch is still only 32 bits.  They said they'd be releasing a 64 bit version in the near future.

Barry

[Jon Hodgson]

Barry Hufker wrote on Mon, 10 July 2006 20:25

I showed the results of this test to the folk at Audio Ease.  They said the Izotope was operating at 64 bits (as the name plainly shows in the test) and Barbabatch is still only 32 bits.  They said they'd be releasing a 64 bit version in the near future. Barry

Actually barbabatch looks very good, even at 32 bits, errors 160 dB below full scale are not going to be audible.

[blue2blue]

Am I misinterpreting those results or does r8brain Free "beat" r8brain Pro in their test?

[Jon Hodgson]

blue2blue wrote on Mon, 10 July 2006 21:57

Am I misinterpreting those results or does r8brain Free "beat" r8brain Pro in their test?

That depends on what you are looking at.

With the setting selected, it is true that pro shows more artifacts (though still below audibility in normal circumstances), but it also allows more high frequencies through. So better bandwidth at the expense of slightly higher noise.

I wouldn't be surprised to find that the pro can be set to give exactly the same results as the free, since that would be a typical approach, use the same algorithm just reduce the options available.

[blue2blue]

DOH!

I'm embarrassed to admit I just realized I hadn't been looking at all the plots. For some reason I stopped when I tried the (disabled) Overall Response plots and had only been looking at the 1 kHz tone...

Wotta dummy I am.   I'll go and look at everything...


Uh, yeah... amazing what a little more info can do...
Never mind.

[Pingu]

blue2blue wrote on Tue, 11 July 2006 04:57

Am I misinterpreting those results or does r8brain Free "beat" r8brain Pro in their test?

That's what i thought too.

I have done a few tests and pro always seems truer to the source for me.

Alexey feels the tests are not accurate.

He also states in the forum that pro is definitely of higher quality.

[blue2blue]

Pingu, make sure you compare the "passband" and "transition" views. (I mean from SRC to SRC, of course.)

I didn't get that interface at first but now that I know my way around it, it seems pretty useful. Finally a useful role for Flash... that seems epochal, somehow. You know, Flash comes of age, or something.

[UnderTow]

Jon Hodgson wrote on Mon, 10 July 2006 17:04

blue2blue wrote on Mon, 10 July 2006 16:10 And on the practical (though subjective) side, your final paragraph's proviso should give some reassurance to those who need to service multiple conflicting formats, ie, video at 96 kHz and conventional CD audio at 44.1 kHz. The most important thing is to get a good SRC. I was quite surprised when I saw this page http://src.infinitewave.ca/ Now I can't confirm the quality of their testing, but if they've done things correctly then it seems to tell us three things 1) It is possible to achieve extremely high quality (170dB is 28 bits worth) 2) Plugins and programs aimed specifically at mastering tend to have higher quality than your standard DAW SRC (different priority, in a multitrack DAW you are trying to keep processor use down so you can have as many tracks and plugins as possible, in a mastering program you can throw all your resources at just two tracks). 3) Even in the worst case in these tests the artifacts are only just edging up into audibility, so use those CPU draining SRCs when you're mastering, by all means, but don't lose sleep if you don't have one.

Just a small comment here: I contributed the Sonar 5 entry. It isn't a realtime SRC. It only happens when you import or export files so processing load isn't part of the equation. It just isn't a very good SRC. No excuses.

This has been mentioned to Cakewalk so they might do something about it in Sonar 6. Especially considering that they allready license stuff from Voxengo. (Like Pristine Space convolution reverb).

Alistair