5.3.2.13089 Public Beta Spdif optimisation?

    Been debating as to whether or not to switch over to 48kHz for many years now, Sam. Made the decision to do so earlier this year, but at this point it's inconsequential 'cause I haven't (and won't be for quite a while) recorded anything yet.


    I figured the minimal disk-space and CPU hit would be more than justified by the high-end gains. Until then I'd figured 44/24 would serve me forever. Now, I'm not-so-sure.


    Interested in your thoughts, bud.

    Quote from hjscheffler

    the only reason to switch to 48 khz is if you are working in the video/soundtrack business. Sonically there is no benefit.

    While I understand what you mean and agree on the lack of sonic benefit, I still have to ask for some more differentiation. :)


    If you setup a recording environment (aka studio), you easily run across plenty of hardware that has been made for 48kHz sample rate (doesn't support 44.1). From a Korg Kronos keyboard or the AxeFX to high end DigiCo consoles.


    Yes, there's probably more digital gear that either sports configurable sample rate or just the simple 44.1kHz S/PDIF. And there's also plenty of gear with no digital interface at all (analog IO). And yes, we can also use analog IO on the Profiler.


    I even agree that it's not Kemper's fault that the digital audio market is still cluttered with a crazy amount of different interfaces and protocols (MADI, Dante, AES3, AES50, (coax.) S/PDIF, (opt.) ADAT, ...)


    And even worse, the OS manufacturers (i.e. Microsoft and Apple) still haven't managed to provide a rock solid and foolproof solution to merge all kinds of USB/Firewire/Thunderbolt/PCIe/AoE interfaces into one big virtual audio interface.


    Bottomline: It's still a nightmare to try to setup an all-digital environment, after all these years. And we always need to find ways to deal with it. Even the smallest help from the manufacturers often is highly welcome. And that's where Kemper comes into play again. External sync and selectable sample rate (at least 44.1 and 48) would have been such a big help and I hope that Kemper will address this in a future product. This is not a situation where Kemper can just say "We're right and there's no benefit of doing otherwise". ;)

    You are right that digital audio connections are a mess.


    But I believe it cannot be the task of makers of hardware music instruments (in detail digital keyboards or digital guitar amps)to solve this problem by providing slave capabilities anddigital signals of all frequencies.
    There is multiple reasons:


    - Hardware music instruments are digitally connected merely by a minority, the majority is mostly used on stage.
    - Designing DSP algorithms for different sample rates increases developement costs.
    - Running DSP algorithms at double sample rate cuts the available calculation power by two, thus reducing the available features.
    - Avoiding the above would require a sample rate converter would again add expenses, that is payed by all, but only utilized by a minority.
    - Instruments with output only (such as keyboards/synthesizers) would still require a wordclock input to be slaved to the audio interface.
    - The user had to be teached to use the wordclock connection for proper synchronisation. I am afraid that only a minority of users are aware of this fact.
    - Correct settings have to be made on both the hardware and the audio interface. Wrong settings are not necessarily identified easily.


    Btw: To my knowledge there is no single digital keyboard or digital guitar amp available with slave capabilities!



    Instead, I think it should be the task of the audio interface makers to provide sample rate converters on their digital IOs.


    - Connecting audio equipment is the main task of audio interfaces, The expense of sample rate converters are where they belong. (compare it to the microphone phantom power, that is provided by audio interfaces as well)
    - Modern sample rate converters are virtually lossless (even though taught differently by many)
    - Several devices could run on different but advantageous sample rates.
    - No wordclock input and cable is required


    - No settings have to be made.
    - The user would use digital connections just as he learned to use analog connections, error save.


    This would be a paradigm change, that can only be pushed along by the makers of audio interfaces.
    RME is making progress actually.
    Spread the word!

    Thank you Christoph!


    Quote from hjscheffler

    the only reason to switch to 48 khz is if you are working in the video/soundtrack business. Sonically there is no benefit.

    This has been my thinking all along, HJ. I suppose I was just second-guessing myself after all this time. Thank you!

    Quote from sambrox

    Wholeheartedly agree.

    Thank you, esteemed SamBro'!

    Quote from ckemper

    You are right that digital audio connections are a mess.


    But I believe it cannot be the task of makers of hardware music instruments... to solve this problem by providing slave capabilities and digital signals of all frequencies.

    Thanks Christoph for chiming in and again I agree to some extent to what you write. :)
    You can't solve the general issue ... but you certainly could make life easier for those who would like to make use of digital IO. External Sync is no solution but a major help for the time being.


    Quote from ckemper

    Btw: To my knowledge there is no single digital keyboard or digital guitar amp available with slave capabilities

    The Korg Kronos is capable to be slave (but not switchable between 44.1 and 48kHz ... 48kHz only)
    The AxeFX is capable to be slave (but not switchable between 44.1 and 48kHz ... 48kHz only)


    Quote from ckemper

    Instead, I think it should be the task of the audio interface makers to provide sample rate converters on their digital IOs ...

    From this statement down I wholeheartedly agree with you. :)
    That's nothing you can solve unless you want to be the first to bring such a product to market, haha.
    But you can certainly be of help by providing the most popular samplerates (44.1 and 48) and an option for slave in a future product. That's all I'm hoping for because I highly doubt there will be a proper audio interface with plenty of digital IO and ASC built-in any time soon.


    Cheers
    Martin


    PS: There was a promising attempt to tackle digital IO integration by TC Electronics called the Digital Konnekt x32. Sadly it was discontinued and not followed by an even more complete (daisy-chainable) solution

    I must admit to regularly using the XLR output on my Kemper, as many projects come through the studio at varying sample rates. With guitar frequencies not occupying much below 100hz or above 10k, it has never been a problem with a decent converter. The outputs on the Kemper are good enough.

    Karl


    Kemper Rack OS 9.0.5 - Mac OS X 12.6.7

    Quote from Netheravon

    Well currently when I change my input from front to SPDIF for revamping, it still takes the input from the front. I'll have to update to this be a to see if it sorts that out. I don't like using Betas tho as I use my Kemper regularly live....

    Two times happens to me too the same problem. I switched to Reamp like I did always with active SPDIF signal but the KPA remains on Front Input.


    There is something wrong sometimes for detect SPDIF connection, I'm opening a ticket for this.

    Actually there is a benefit to 48k or above even. The issue is that reconstructing frequencies up to half the sample rate actually only applies to infinite series, meaning if you had the same sine wave at anything below 22.05k (but not including) you can accurately determine its frequency, magnitude and phase from those samples... with an infinite set of samples. This is of course self evident. The issue is accurately reconstructing a frequency from a less than infinite sample set is not so trivial, realistically you're looking at rather lower actual reconstruction floor than half for accuracy, you really need at least four samples before you have true near accuracy for a single cycle so only a quarter of the total frequency.


    Fortunately enough people's ears are really only sensitive in the mids, where voices are and the sounds of our various predators and prey, so it doesn't matter too much most of the time. Especially for basic playbakBut it rears its head once you start to process the signal. If you start to generate square waves from e.g. Distortion or compression effects you will start to get so called aliasing artifacts, that's sub-harmonics caused by the attempt of the aliasing filter which is typically using a form of sinc filtering, trying to resolve every single frequency (which is what Fourier tells us a true square wave must contain), this is especially apparent when it's a modification of another existing oseudo chaotic waveform such as a guitar. Basically you get these fringes if you apply sinc to a highly contrasting image, same deal with a 1D dataset such as an audio wave. So the solution to this is to do your processing at a higher frequency. This reduces the aliasing artifacts substantially.


    These days most plugins do in fact upsample before processing, just like the Kemper does but not all do. Using a higher rate ensures that you do not encounter these issues. Within certain industries as well there are standards which are not 44.1k, keeping to the standard throughout the pipeline is considersbly less effort all round (especially if you ever have to deal with little bespoke tools people have flung together than would simply treat a 44.1k file as if it were 48k and play it back sped up!), so enforcing your own views on what should or shouldn't be used is both counterproductive and likely to leave you with fewer gigs.

    Quote from Per

    Actually there is a benefit to 48k or above even. The issue is that reconstructing frequencies up to half the sample rate actually only applies to infinite series, meaning if you had the same sine wave at anything below 22.05k (but not including) you can accurately determine its frequency, magnitude and phase from those samples... with an infinite set of samples. This is of course self evident. The issue is accurately reconstructing a frequency from a less than infinite sample set is not so trivial, realistically you're looking at rather lower actual reconstruction floor than half for accuracy, you really need at least four samples before you have true near accuracy for a single cycle so only a quarter of the total frequency.

    The thing is, a quarter of 44100 Hz is 11025. The first harmonic of a square wave is an octave above, which would be 22050 Hz. As humans can only (theoretically, at an absolute maximum) hear up to 20000 Hz, we can't distinguish between a sine wave and a square wave above (again, at a theoretical absolute maximum) about 10000 Hz. Therefore, it's pretty irrelevant. This video might be interesting to you :


    External Content youtu.be
    Content embedded from external sources will not be displayed without your consent.
    Through the activation of external content, you agree that personal data may be transferred to third party platforms. We have provided more information on this in our privacy policy.


    As guitars and pretty much everything apart from pure, synthesised sounds have much more rich and complex harmonics that determine the timbre, digital audio can actually reproduce these real-world sounds much easier than pure waves (though it also doesn't have any difficulty reproducing synthesised sounds within the range of human hearing, either). As for some plugins creating aliasing artefacts in the manner you described, I think it's been a generation or two since that was the case, at least in my experience.

    Edited 2 times, last by sambrox ().

    Quote from sambrox

    The thing is, a quarter of 44100 Hz is 11025. The first harmonic of a square wave is an octave above, which would be 22050 Hz. As humans can only (theoretically, at an absolute maximum) hear up to 20000 Hz, we can't distinguish between a sine wave and a square wave above (again, at a theoretical absolute maximum) about 10000 Hz. Therefore, it's pretty irrelevant. This video might be interesting to you :

    External Content youtu.be
    Content embedded from external sources will not be displayed without your consent.
    Through the activation of external content, you agree that personal data may be transferred to third party platforms. We have provided more information on this in our privacy policy.


    As guitars and pretty much everything apart from pure, synthesised sounds have much more rich and complex harmonics that determine the timbre, digital audio can actually reproduce these real-world sounds much easier than pure waves (though it also doesn't have any difficulty reproducing synthesised sounds within the range of human hearing, either). As for some plugins creating aliasing artefacts in the manner you described, I think it's been a generation or two since that was the case, at least in my experience.


    That would only make sense if the output of 44.1k -1.0, 1.0, -1.0, 1.0... (perfectly phase aligned peaks and troughs to the samples) were to actually be a square wave. But it wouldn't be, it would be a sinusoidal wave through any meaningful audio reconstruction filter, you're trying to resolve a moire pattern into it's core original features. Simply put - you can't get back information that's been removed or aliased out, just like you can't de-blur a 2 pixel image of a car and get it's number plate in glorious 4k detail. Sorry CSI.


    Even with the best AI driven reconstruction systems you will inherently lose data with lower frequency discrete sampling methods. The amount of data lost matches the basic inverse square law (that old chestnut rears it's head once more) as frequency is reduced, when you try to upsample to get to your original dataset there will be increasing discrepancies across the whole range.


    If you use a filter specifically designed for a specific waveform then assuming that waveform was the source you can retrieve it with sufficient samples at anywhere up to Nyquists observation of under half the frequency. If you are not aware of the source waveform then you can still retrieve it if you have infinite samples. If you have neither then you are limited to approximations which will likely be based on sinusoidal forms (and that's all you need for audio).


    I'd like to emphasize again that due to the nature of our ears and brains, and the quality of modern reconstruction filters 44.1k is perfectly adequate for straight playback. The only audible issues occur during processing with the ever dwindling collections of plugins (and DAW's) that don't upsample prior to processing. You guys may not have been here for this, or forgot about it perhaps, but the Kemper had it's own aliasing artifacts issues earlier on, they were resolved in a firmware update, but it was a big deal here on the forum.


    Quote from Michael_dk

    What do you mean by "less than infinite sample set"?

    Any series that isn't infinite in length. i.e. every real world signal where high frequency signals simply die out quicker with less "energy" than low frequency, or where signals are chaotic in nature with many different frequencies coming and going throughout the duration of only a relatively few samples, most of that super treble stuff we get comes in the transients for example.


    Most reconstruction filters only deal with a tiny window into the sample set too, so you basically get a smoothed out version of the waveform as directly represented by it's samples with a little overshoot rather than an attempt to really analyze and reconstruct the true waveform. By tiny I mean only a handful of samples, likely a lot less than your input buffer in your DAW.


    The other factor in all this is that the quality of that top end is also highly contingent on the sampling going in. Record something like Cymbals with an interface that doesn't have a good word clock and really good lo-pass filter and you end up with basically just static swish white noise. The lo-pass is particularly critical as there are of course frequencies above audible but when you sample discrete points in time they can have enough energy to create harmonic effects and just straight up noise in the sample set. Then the word clock jitter is of course actually better resolved at lower sampling frequencies, so it's one argument against raising the sampling rate (much debated on Gearslutz).


    Personally speaking I record at 48k and my audio interface has a nice feature which allows it to upsample SPDIF to whatever you're using. It might seem like that's doing nothing but raising the file size, but it actually does improve matters just enough. It has no impact on the Kemper output of course, but I use Reason with a lot of effects and processing and not all RE's upsample, so 48k is an OK compromise, it reduces aliasing issues to the point of not really being annoyingly audible, doesn't introduce jitter problems and the file sizes don't become ridiculously huge. Then I export to 44.1k (unless there's good reason not to).

    Quote from Per

    Any series that isn't infinite in length. i.e. every real world signal where high frequency signals simply die out quicker with less "energy" than low frequency, or where signals are chaotic in nature with many different frequencies coming and going throughout the duration of only a relatively few samples, most of that super treble stuff we get comes in the transients for example.

    I don't think I understand. Yes, you only have two samples per period at the nyquist frequency, but you don't have any frequencies above that (due to the low pass filter) - the nyquist frequency is the absolute worst case. All the other frequencies get more samples per period than that one.

    Quote from Michael_dk

    I don't think I understand. Yes, you only have two samples per period at the nyquist frequency, but you don't have any frequencies above that (due to the low pass filter) - the nyquist frequency is the absolute worst case. All the other frequencies get more samples per period than that one.

    Which is still not enough. Nyquist doesn't just describe the worst possible case, it's an observation of the patently obvious which is that you can with an infinite sample set accurately recreate every single frequency up to (but not including) half the sample rate.


    The critical part of that is "with an infinite sample set", which is simply not realistic with real world audio.


    In the real world reconstruction algorithms use a very small sample set and a basic bastardized sinc filter (most common ideal, however again requires infinite set so in the real world has a smaller set and is correspondingly less accurate) for the job of calculating what the original waveform should have been. You need a frequency of at least a quarter to be able to recreate the original source with any real accuracy over a tiny sample set.


    Think of it this way, lets say you have a frequency that's anywhere between 22.05kHz and 11.025kHz using 44.1k as the overall sample rate. What will occur is moire patterns in the sample set, phase and frequency mean you cannot possibly have "ideal" peak/trough samples throughout the set. This means that with a small sample set the accuracy is going to go out of the window. Just use your imagination to visualize the sample set and how a real world reconstruction filter utilizing a small sample set cannot possibly recreate these frequencies accurately either in magnitude, phase or even frequency.


    There's no hard cutoff where suddenly recreaction becomes accurate at exactly half of the sampling frequency except with an infinite sample set (of the same frequencies not changing in any way). Nyquist's observation isn't that every frequency below half the sample rate can be accurately determined and recreated with a finite sample set.


    Instead when dealing with finite sample sets there is a gradual but logarithmic degradation of accuracy of reconstruction throughout the frequencies as you go higher, till you reach half the overall frequency, at which point with any number of samples or reconstruction window-size, accurate reconstruction is no longer possible. The longer the frequencies stay put and the larger the reconstruction window, the greater the bias towards accuracy within the frequency domain will be.


    Fortunately enough (and to reiterate myself) in the real world accuracy in the higher frequencies isn't that critical, psychoaccoustics tells us we're not very good at discerning pitch above 11kHz anyway, above that we discern pitch more through effects like interference, so once again 44.1k is perfectly fine for audio playback... for humans.

    Here, maybe this will help clarify (second half) :

    External Content www.youtube.com
    Content embedded from external sources will not be displayed without your consent.
    Through the activation of external content, you agree that personal data may be transferred to third party platforms. We have provided more information on this in our privacy policy.


    With a static waveform then, no you don't need peaks and troughs, with sufficient samples you can determine both frequency and amplitude correctly.


    In the real world though signals are complex, they contain many frequencies at many magnitudes with many durations.


    Without having the assumption of a static single frequency signal or infinite sample set it actually becomes impossible to determine the original waveform with great accuracy. It's literally the CSI example. You're asking for a quart out of a pint pot.


    The trouble with Nyquist is that it's talking about theoretical signal reconstruction. In the real world of audio signal processing it becomes less about "you can accurately determine and reconstruct a waveform of frequency under half the sampling rate" and more about "you cannot accurately determine and reconstruct a waveform of frequency above or equal to half the sampling rate". These two statements are not the same thing, yet people somehow imagine that they are, and that Nyquist means everything under half the sampling rate is accurately reproduced. That's simply not true. It's mathematically impossible because with complex signals and short sample sets there may in fact be multiple solutions.


    During reconstruction all that you can really do is generate a waveform that fits through the samples you have and hope for the best, the beauty of the sinc function is that even with a restricted range it helps iron out some (but not all) of the magnitude issues at that upper end, the downside is that it introduces it's own artifacts to either side of sudden changes in value.