Paul Frindle intersample overs

[wsaidah]

Paul Frindle intersample overs does any one no were to get thit paper

[Kiwburger]

http://www.google.co.nz/search?hl=en&q= ... arch&meta=

[Kiwburger]

I've read a lot of Paul's stuff - it's spread over many forums. The main idea that I got was that it's important to realise with digital audio - what you see is not what you hear. The analogy that is often made, comparing digital audio with digitised pictures, is very misleading. With pictures, you see every pixel, and therefore every bit counts.

With audio, the bits are converted into a continuously varying analog voltage. You never "hear" pixilated audio - it is always seamlessly converted to a varying voltage. That might seem obvious - but there are some misconceptions that exist that Paul attempts to blow away.

The idea that using all the resolution of a converter is necessary or good - only leads to distortion.

Two things worth knowing: you can have digital samples that don't clip, and yet the converted analog waveform can clip. Also - you can have bits that are never converted literally - because the conversion process sort of averages out some of the information.

The idea that low level waveforms loose resolution is just wrong. This is because all audio is composed of sinewaves. You can define a sinewave with only two samples. There is only one sinewave that can "connect the dots", so that's what your converter does. It doesn't matter if you use 2 samples or 2000 samples - the converter will still re-create the same sinewave. As long as the sample rate is fast enough, you will get all the frequencies/harmonics of the sound. So worrying needlessly about "preserving bits" just causes you to track too hot and get nasty distortion.

I've found that certain plugins clip waay under 0dBFs, and they are a much greater problem imo. I used to use sinewaves to test digital chains - but i've found a better, more revealing way. Use the frequency swept sinewaves that are used for deconvolving impulses. They will reveal some really nasty weaknesses in many plugins, which surprised the heck out of me. No need to worry about possible intersample clips if you are running throug a plugin that clips internally at around -20dBFS ...

[kylen]

That's interesting Kiwi - the part about how the A/D/A kind of takes 'artistic license' with the samples or whatever. I'll bet Aleksey has probably thought about this with his R8Brain Pro SRC. Both digital and analog reproduction have weaknesses brought about by whatever is in the signal chain. But after running a couple of digital EQs or other plugins across your audio I wonder how long the pristine quality of great converters is preserved anyway...

[defjamm]

I've found that certain plugins clip waay under 0dBFs, and they are a much greater problem imo. I used to use sinewaves to test digital chains - but i've found a better, more revealing way. Use the frequency swept sinewaves that are used for deconvolving impulses. They will reveal some really nasty weaknesses in many plugins, which surprised the heck out of me. No need to worry about possible intersample clips if you are running throug a plugin that clips internally at around -20dBFS ...

could you pm me which ones have this problem and describe the exact testing method?

[Aleksey Vaneev]

The problem when fighting the intersample overs is the absence of knowledge on how the user end reproduction D/A converter behaves. You may suppress intersample overs with Voxengo Elephant limiter, for example (at 4x oversampling). But this will be only applicable to linear-phase DAC designs that perform oversampling before passing the digital stream to D/A converter. Moreover, I do think that such designs have 1 bit more of headroom (17 bits for 16 bit output), so that intersample overs are handled correctly in any case.

If no such oversampling is used, and the DAC is of PWM type working at the same sample rate as the original signal is stored, and applies a steep analog low-pass filter after the DAC output, to suppress spectral mirrors, the only problem will be the resolution of analog electronics following the DAC. I think in most cases such electronics have enough headroom to handle peaks and output line level signal without any distortions.

I guess, D/A converter manufacturers should publicize information about their converters' signal paths. Otherwise no stable solution can be found.

[Zaphod (giancarlo)]

..simply keep the general gain for each track below -6dBfs.

Paul says also a different thing: a lot of fft-based equalizers "reconstruct" the signal. And they do in a bad way if the level is too high (close to 0dB). So keep your track below -6dB, every plug will work in a better way even if it doesn't display an overflow error.

@defjamm: someone displays errors using a different algorithm.

[Aleksey Vaneev]

Paul says also a different thing: a lot of fft-based equalizers "reconstruct" the signal. And they do in a bad way if the level is too high (close to 0dB). So keep your track below -6dB, every plug will work in a better way even if it doesn't display an overflow error.

I do not understand this point at all. Was this said in the context of plug-ins or D/A converters? Because if we are talking about plug-ins, there is no such thing as 'reconstructing' the signal, and there is no such thing as 'fft-based equalizer', because FFT is just a method of performing convolution, it is not a 'base' for equalization processing (like it is in some fundamental way is different from the 'brute' convolution). Not to say, floating point processing practically has an unlimited headroom - it is absolutely transparent whether you are processing -60 dBFS signal or 60 dBFS signal, in the case of equalization.

[wsaidah]

so no link to the paper

[Zaphod (giancarlo)]

I do not understand this point at all.

Ok, I'm quoting an old topic from REP about this topic.The author is... Paul frindle (known as Sony Oxford developer and old SSL eqs developer)

Exactly Paul...that's why I included:

"...while of course cross referencing the peak DAW meters as well."

Thanks!



Actually (at the risk of putting the cat amongst the pigeons) I can suggest a simple experiment people can do themselves to illustrate this in action in the most graphic way, which should dispel any lingering doubt that it's important.

The aim is to show that what looks like a legal 'signal' way below any red light in your system can still represent something that cannot pass even remotely correctly out of your digital mixer at full level. And also to illustrate how this may affect your sound quality in practice when mixing ITB. It's a kind of worst case scenario - but it illustrates the problem.

You need a W/S like ProTools, a signal generator plug-in that has a good filter section that actually goes flat to 20KHz and rolls off at 24dB/oct or so.

- In Pro tools get a mono channel up,

- stick the PT generator plug-in at the beginning of the channel and set it for sine at say 1-2KHz.

- Follow this with a good filter plug-in set for the max slope at 20KHz. (For example the Oxford EQ plug-in has 36dB/oct at 20KHz and illustrates this well - any other good HF filter should work as well).

- As an initial test set the channel fader at 0dB and note that the PT meters shows the sinewave signal at -6dBr and that putting the filter plug-in in and out using bypass has no effect.

- OK now switch the signal genny to white noise and note that the level on PT is still -6dBr.

- Now un-bypass the filter plug-in and watch the signal level rise dramatically!! In the case of the Oxford 36dB/oct filter the meter level will rise a full 5dBr to nearly flat out.

Ok so what's happening - how is this possible? Well the digital genny plug-in produces sinewaves correctly - but when in noise setting it is just a random number generator driving the output. So although when set to -6dB peak value no sample ever gets to be greater than 50% modulation - a reconstruction of the undecoded SAMPLE VALUES produces nearly full level SIGNAL. Reconstruction means filtering and so the filter plug-in is acting like a partial reconstruction filter (much like a DAC) - which in turn is now feeding a more legitimate SIGNAL which the sample value only meter can read more correctly.

Ok now if this SAMPLE train is passing out of your DAC it too is being reconstructed correctly - so this -6dBr noise from the genny would a produce nearly full modulation SIGNAL if you fed this to the DAC directly - even though no sample gets to be bigger than 50% and no reading say's it's bigger than -6dBr.

If your filter is a good one you should be able to switch it in and out and hear no difference in the sound of the signal from your DAC - despite the PT meter reading wildly different. The filter has neither added nor taken anything significant out of the intended audio signal - but you have nearly doubled the sample values within the PT channel!

Ok, now wind the genny level up to say -2 or -3dB (still less than only 75% full level) and do the same thing. What happens? Well it now clips when the filter is in (samples bigger than flat out) - now the sound definitely changes when you switch the filter in and out - because it is mathematically limited and in error when the filter is in - cos it cannot pass through TDM slot at the output of the filter!!

That is what would be happening in your DAC, it would saturate if you sent this at only 3dB setting on the genny - reading -3dBr within the mixer itself, straight to the output!!

Ok now what does this mean for a mix? Well with all those mixed signals, cymbal crashes, HF EQ and limiting etc.. how close do you imagine the output signal can get to being a bit like white noise in places within a real production - even if none of the contributing channels hit the red light? Is this not the exact register of what we term as 'air' and 'resolution'? And people are aiming at max possible mix output levels on meters that do not show SIGNAL.

So why does an OTB mixer apparently sound better than an ITB mixer when you are modulating your digits close to 0dBr (sample value) all over the place? Well all those DACs (flawed as they may be) are acting to legitimately reconstruct your programme - before - you mix them all together and produce too many illegal signals that cannot pass out of your digital mixer! Paradoxically, the loss of sound quality due to all those converters is not as bad as the illegal signals created within the digital mixer by the 'too hot' signals you are trying in vain to pass out of the system.

It is not a summing issue at all (the one thing digits CAN do is add up almost perfectly). It's an illegal output problem caused by the fact that there are no meters that display actual SIGNAL in your whole mixing environment - you simply never see it happening.

So - go back and get your fav test mix back up on your W/S, re-mix the whole thing making sure that at every place in all chains (including between all plug-ins) never gets bigger than -6dBr. Make sure your final output after any limiting etc also never peaks beyond -6dBr. Now do the comparison between this ITB mix and a similar OTB mix. You might have a big surprise

excuse me.. it's a bit long
This thread is legendary in R/E/P

[Zaphod (giancarlo)]

And again (here he is speaking about equalizers)

Yes indeed this is so - anything that 're-arranges' the phase and response can get you HIGHER peaks - in fact up 6dB higher for LF roll-off of an already totally clipped signal - the need for headroom isn't just intersample peaks, it's your whole freedom of artistic expression within your mix that's at risk.

For instance - if the mastering engineer decides to roll-off some excess LF from your programme he could well end up having to REDUCE it's level (and all-important loudness) to accomodate the extra peak values!!

In the experiment I described you can try this too by taking out the HF filtering and putting in LF filtering instead - and watch the levels rise
Another interesting one is to get a squarewave from the genny at say 200Hz -6dB and then insert a filter set to cut off below 100Hz - the peak level will rise 6dB (or more if the filter is a high order). Ok this is an extreme example (not very musical) but if this was your programme at full level then it would have to be represented at 6dB (or more) lower than the original in order to avoid actual clipping!! Food for thought for those that like to clip the kick drum a teensy bit for that little bit extra attack and presence. And of course the same thing would apply at higher freqs (such as vocals) if you let it saturate a bit for effect then EQ it a bit to roll-of some of the 'raspy edges'.

This is not new stuff - and the same things happen in analogue processing as well - it's just that analogue systems have signal operating levels 10's of dB below signal clipping - and analogue tape recording methods were more tolerant cos they could accomodate significant overload at LF (as much as 10dB) and saturated softly at HF (i.e. produced that fuzzy splashy 'air' in the presence of HF overload - rather than hard and ear-grating clips).

[Aleksey Vaneev]

That is what would be happening in your DAC, it would saturate if you sent this at only 3dB setting on the genny - reading -3dBr within the mixer itself, straight to the output!!

I think this is not a universal conclusion. If ProTools DAC clips intersample overshoots - it is the problem of ProTools, not of DACs. DACs can be designed in various ways - including ways where intersample overshoots can be represented correctly in analog form, even if the original digital track peaks at 0dBFS all the way.

[Zaphod (giancarlo)]

I agree.
Paul is aware of it. It isn't an universal conclusion.
Here the forum is very professional and most users have pro-tools. Anyway the message is: you don't know if the plug is reporting an overflow in the right way, you don't know if your dacs are repesenting well overshoots. If you know, trust in your meters. If you don't know, keep levels a bit lower. I should say this is working on my side.