advertisement


Upsampling--Good or Bad?

stackowax

pfm Member
I'm getting a jitter reducer/upsampling device made by Monarchy Audio (mainly so I can connect one device that only has optical out [a DSPeaker Anti-mode 2.0 Dual Core] to another device that has only a coaxial in [Resolution Audio Opus 21]).

But it has occurred to me, after the fact (it hasn't arrived yet), that maybe there are downsides to upsampling? It seems like, in audio, there are always downsides.

FWIW, Monarchy makes other similar devices and I couldn't really figure out why one would be preferred to another. But Monarchy thought the one I've linked to above would be best, so that's what I'm getting.
 
In principle any avoidable resampling process risks losing or altering some detail or accuracy from the data. But in practice if the 'downstream' device isn't perfect then it might avoid or reduce some problem elsewhere in the rest of the system. So the short answer to your question is, "It depends." 8-]

Jim
 
My Cambridge 851N up samples everything and it certainly doesn't sound bad. You can't turn the up sampling off so there's no way to do an A/B test but I doubt there'd be a significant difference either way. It does make low-res 64, 96 and 128k streams sound pretty good though.
 
Just a note on how upsampling and oversampling differ:

Upsampling involves increasing the sample rate of a digital signal from its base rate (eg 44.1KHz for Red Book CDs) by any selected factor which need not be an integer value.

Oversampling is a limited subset of upsampling where the factor values are limited to integer values.

So, upsampling can be done using non-integer factors such as 1.3333 or 2.5 etc. while oversampling uses factors that are integers such as 4x or 8x.

Using the Red Book standard for CDs, an upsampling at 1.3333x will result in a sample rate of 44.1 x 1.3333 or 58.8 KHz, and oversampling at 4x will result in a sample rate of 44.1 x 4 or 176.4 KHz.

In a Non-Oversampling (NOS) DAC, the digital signal's sample rate is not raised and is fed to the D/A converter which converts to an analogue signal, but the D/A conversion process introduces "aliases" (at multiples of sample rate) which, for optimal SQ, need to be filtered out via the use of anti-aliasing filters which, have to be implemented in the analogue stage (as the signal is already an analogue signal after D/A conversion. These analogue anti-aliasing filters are complex circuits and are difficult to implement.

This difficulty is introduced by the need to provide three filter types that all need to work together:

a) A pass filter for signals with frequencies below 20 KHz (the audio signal)
b) A stop filter for signals with frequencies above 24 KHz (any aliases)
c) A transition filter covering the 20KHz to 24KHz "gap" that transitions from for the 20KHz requirement to "pass" to the 24KHz requirement to "stop" (or in gap terms needs to attenuate at a steep slope of "pass" (zero attenuation) at 20 KHz to "stop" (infinite attenuation) at 24KHz - a small frequency band of just 4KHz

This transition filter is where the complexity arises and the steep slope has resulted in the use of the term "brick wall filter".

The concepts of upsampling (non-integer sample rate multiple) and oversampling (integer sample rate multiple) were introduced to widen the gap between the pass and stop filters in order to make the slope of the transition filter much less steep (and, thereby, less complex to design and build).

The two basic options (if we exclude upsampling) are, therefore, oversampling (or OS) DACs and non-oversampling (NOS) DACs.

Decent NOS DACs with effective analogue anti-aliasing filters tend to be costly due to the complexity and component costs associated with such filters.

OS DACs can be built at a lower cost as less-steep filters can be built for less.

NOTE: I've attempted to simplify the above and, as a result, may have strayed into over-simplification (as opposed to up-simplification - sorry, couldn't resist... :) ) - so if anyone needs to clarify or expand on this, please feel free.

Dave
 
The snag here isn't oversimplication. It is that people tend to use the terms in different ways that can overlap or clash.

I've tended to use "oversample" to indicate that the sample rate is higher that necessary from an Information Theory POV. e,g, in an ADC that samples and runs at a rate well above 48k but takes its input from an analog signal with a nominal bandwidth that isn't above 24k and finally outputs 48k.

But I have also at times used Oversample to mean a simple 1:integer ratio and Upsample to mean ratios that don't meet that criterion.

A similar snag with 'NOS' is that in some cases it seems to assume you then don't bother to fully employ a following analogue filter to limit the output according to Nyquist. But in other cases it does. Thus it may be putting the emphasis in the wrong place as a description of what people are actually focussed upon.

Then, of course, we have 'resampling' which may be a conversion with the *same* nominal input and output rates, but with no fixed synch. Or may cover some of the above! :) The BBC tend to have asynch resamplers about the place. Saves them having to phase lock their entire operation to one clock.
 
SQ wise - a matter of taste, IME. i have been trying this recently, up to DSD 256, first impressions are good but i'm still on the fence over DSD as for some reason, it doesn't seem too smooth that comes with up sampling i have done before (not to DSD) and does do something i quite like. struggling to put my finger on or express what seems good about it but it warrants time until i arrive at my own conclusion. Mojo and certainly new Chord DAC owners should be aware this isn't something you should do as Mojo, hugo et al over sample with much, much more power than any home computer.

back to topic, up sampling (excluding to DSD) is too smooth for my liking. takes away dynamics. i tried it for a while, but when i reverted back to native, back came the sound i love and get drawn into. again, it is a matter of taste. i don't think there is a right or wrong.

i look forward to up sampling more to DSD with Audirvana however you do need a powerful computer for DSD up sampling. my audio machine is a trash can mac pro, 6 core, 64GB ram so no slouch yet any up sampling above DSD64 gives me drop outs, despite 50% CPU usage, i still get drop outs so i think i need to change the settings which seems to be the done thing. no signs of over heating at all in my case, this is a common finding but this could be down to the mac pro's cooling system which is superb. having read online, contrary to what you might think, when using Audirvana and up sampling to DSD it is recommended to drop RAM buffer down to 1.5MB! this seems strange but maybe so the CPU is doing computing on the fly, rather than having to do an awful lot more at once before it goes into the RAM itself. i may be wrong but lowering the ram buffer does help drop outs. there is an interesting blog on the Audirvana site about up sampling to DSD and what settings people are using. i have always used the default settings but as above, soon as i started up sampling to DSD, i had to alter several settings as per recommendations on the blog. this helped but i'm still getting drop outs. hopefully i will get a reply from damien about it soon and i can listen properly. i suspect i'll be reverting back to native but it costs nothing to see what it's all about and offers. anyone up sampling to DSD should keep an eye on CPU temperature as a precaution unless you don't mind buying another, assuming you can.
 
But it has occurred to me, after the fact (it hasn't arrived yet), that maybe there are downsides to upsampling? It seems like, in audio, there are always downsides.

Technical downsides galore in audio, the only 2 questions worth considering are:

- Audibility. Is the process, here up-sampling, causing audible effects.
- System interface/stability. Is the process causing unwanted signal to enter the chain and upset anything downstream.

IME audiophiles spend far too much time fussing and worrying about inaudible, but sometimes measurable effects. This is especially true when it comes to digital audio.
 
Upsampling vs Oversampling - my experience (as a DSP programmer) is rather different. I'd say that upsampling was a *process* which was used to produce an oversampled signal.

Basically, if you have a signal which is bandlimited up to frequency f, then the optimal sampling rate to capture this signal is 2f (twice the maximum frequency present in the original signal). Any sampled data for such a signal which has a sampling rate >2f will be adequate, but would be considered 'oversampled' in that there were more samples present that were strictly necessary to represent the original signal.

So, by this definition, we upsample a digitized signal to produce oversampled data. Why would we ever do this? Well, it turns out that if we have some downstream signal processing stages which are non-linear (for sample, say we wanted to simulate the sort of distortion that valves introduce) then we need more bandwidth for the non-linearities to introduce un-aliased frequencies, otherwise we pollute the pass band (the bit we want to listen to) with these aliases. By oversampling, we can have a representation with plenty of spare room for the distortion to sit without aliases.

The process is then something like:

Signal -> oversampler -> non-linear processing -> decimator -> Processed Signal

Basically, we oversample (say 16x) apply our non-linear processing on the oversampled data, then decimate back down to the original sampling rate (lots of ways to do this, outside the scope of this) and hence produce a processed signal without annoying aliases messing up the signal.

In the context of DACs, oversampling is a perfectly fine technique, and looses no information, at the expense of more computation within the DSP code (since we have more samples than we need), but may mean that some non-linear steps can be applied with simpler algorithms, say, simplifying some filter stages, or pushing some noise outside the audible region. There's no hard and fast rule for what is better.

The thing to remember is that oversampling doesn't damage the signal, it can be converted back to the original data in a bit perfect fashion. This is one of the reasons that people are confused by this sort of thing, the maths is hard, and some of the results are counter-intuitive if you come at it from a 'any processing must have a detrimental affect' point of view.
 
The thing to remember is that oversampling doesn't damage the signal, it can be converted back to the original data in a bit perfect fashion. This is one of the reasons that people are confused by this sort of thing, the maths is hard, and some of the results are counter-intuitive if you come at it from a 'any processing must have a detrimental affect' point of view.

Your statement that it "doesn't" is, I think, too absolute an assertion. It can be done in ways and circumstances where no information is lost. So is possible. But that doesn't ensure *all* systems that employ oversampling totally preserve the contained information. Reality doesn't always meet the ideal. The devil will be in the practical details.

Might have been better to say "doesn't always/inevitably".
 
Your statement that it "doesn't" is, I think, too absolute an assertion. It can be done in ways and circumstances where no information is lost. So is possible. But that doesn't ensure *all* systems that employ oversampling totally preserve the contained information. Reality doesn't always meet the ideal. The devil will be in the practical details.

Might have been better to say "doesn't always/inevitably".

That's true, I did wonder about saying that, but it kind of muddies the water - the reality is that it's such a well understood process that it *should* be done perfectly these days, and when I say 'these days' I mean in anything produced in the last 15 years or so. I'd probably use 2000 as the cutoff for DSP CPU being available that means that there is no excuse for using an inferior algorithm based on not having enough CPU performance to do it properly.

The difficult times were when fixed point DSP was king, when we were all writing DSP for motorola 56k processors running 100 mips - that did make some stuff like this tricky to do well, but it was still possible, it was just rarer.

But anyhow, point taken, I shouldn't assume it's done well just because it can be!
 
I REALLY don't understand what's going on with digital audio (although, to be fair, I really don't understand what's going on with analog audio either!).

With that said, here's another question. I've inserted the jitter reducer/upsampler into the (disturbingly) growing chain of digital components I have as follows:

1. Squeezebox Touch>DSpeaker Anti-mode 2.0 Dual Core>MONARCHY 48/96 DIP UPSAMPLER>Resolution Audio Opus 21 Digital Input

Although it would complicate things, would it make more or less sense to do this:

2. Squeezebox Touch>MONARCHY 48/96 DIP UPSAMPLER>DSpeaker Anti-mode 2.0 Dual Core>Optical to Coax converter>Resolution Audio DAC
 
I always maintain that the best sound I ever got from CD was using native 16 bit DAC chips.

I owned the original DCS kit way back when to boot but the 1541 was somehow simply musically right to my ears.

It made sense for Dac manufacturers to move architecture to 24 bit simply because they could offer a product that was suited to universal applications with the advent of DVD. It did however mean asyncronous conversion of native CD data which was a potential weak point.

Having said that, the differences we are talking about with any of this are tiny and verging on insignificant to all but a very few.
 
I REALLY don't understand what's going on with digital audio (although, to be fair, I really don't understand what's going on with analog audio either!).

With that said, here's another question. I've inserted the jitter reducer/upsampler into the (disturbingly) growing chain of digital components I have as follows:

1. Squeezebox Touch>DSpeaker Anti-mode 2.0 Dual Core>MONARCHY 48/96 DIP UPSAMPLER>Resolution Audio Opus 21 Digital Input

Although it would complicate things, would it make more or less sense to do this:

2. Squeezebox Touch>MONARCHY 48/96 DIP UPSAMPLER>DSpeaker Anti-mode 2.0 Dual Core>Optical to Coax converter>Resolution Audio DAC

the antimode outputs 48k only via toslink, so i see no purpose to upsample or mess with the digital signal before it...or after it for that matter...
 
the antimode outputs 48k only via toslink, so i see no purpose to upsample or mess with the digital signal before it...or after it for that matter...

Does it? I know it takes a 96k input via toslink. I assumed it output 96k too.
 


advertisement


Back
Top