what is the difference between non-oversampling and oversampling in a dac?

[RnB180]

whats the difference and which one is better?

what would the micro dac be classified as?

 

[Teerawit]

Personal preference -- some people prefer to leave the signal unchanged, some like it to oversample.

 

[blessingx]

A couple paragraphs into this review gives a brief description.

 

[lan]

The CS4398 DAC chip used in the MicroDAC oversamples.

There's no such thing as which is better. It just depends on which you like more.

 

[kontai69]

Here is a good length thread discussing the subject...
http://db.audioasylum.com/cgi/m.mpl?...ng&r=&session=

 

[JaZZ]

In the beginning of the CD era non-oversampling players were the norm and Philips the only manufacturer who settled for (4x) oversampling in a 14-bit design. But oversampling gained ground because it enabled cheaper production. In the early days oversampling was a mere linear interpolation between the sample points; nowadays it's based on sophisticated algorithms with a sinc function. The latter indicates what oversampling actually is for: not to widen the bandwidth, but quite the opposite: its low-pass (pre-)filter function -- that's why it's also called «digital filter» -- enables easier analog filtering with less electronic components in the signal path. A modern implementation with asynchronous (= non-integer) oversampling is often called «upsampling» and represents an actual sample-rate conversion. Even more so than synchronous oversampling it's said to offer a somewhat «analog» sound or a similarity to hi-rez formats. But one has to consider that whatever upsampling does, it can't magically add more resolution and detail than what's on the CD. And not all listener like the specific sound.

The fundamental opposite of up-/oversampling DACs are filterless (misleadingly often called NOS) designs, to which the above-mentioned Scott Nixon Tube DAC can count, although it actually incorporates a mild analog low-pass filter at 65 kHz with -6 dB point at 120 kHz to minimize ultrasonic noise. The idea behind filterless designs is to prevent the ringing («Gibbs phenomenon») -- a visualization of the notorious time smearing -- resulting from the typical steep low-pass filter necessary to prevent aliasing. The downside is the typical amplitude modulation occurring in the raw, unfiltered signal after D/A conversion -- the result of an interference («beat») between sampling rate and signal frequency, due to the barely more than two samples per wave cycle (see this schematic graph). The amplitude modulation measures as 3-dB drop-off at 20 kHz. The trick of the common «restoration filters» is to introduce a massive filter resonance capable of completely eliminating the amplitude modulation, at the price of the above-mentioned transient smearing. Filterless designs accept the amplitude modulation as well as the risk of a certain amount of aliasing, but considering the anti-aliasing filtering during A/D conversion it's actually minimal.

Now it's up to the listener to decide which flaw he/she prefers.

 

[Scrith]

The graph you linked it interesting...as a software engineer, I look at that unfiltered output and think that surely there must be a way to improve upon that so it more closely matches the original waveform. Do you really mean to tell me that nobody has come up with anything that will create a filtered version of that data (at a higher sampling rate) that is audibly a clear (to anyone who might hear it) improvement over just playing back those unmodified values with a linear interpolation between the points?

 

[JaZZ]

Quote:

Originally Posted by Scrith ...surely there must be a way to improve upon that so it more closely matches the original waveform.

Be aware that no «restoration» algorithm is able to know how the «original waveform» looks like -- it might very well be that it consists of this kind of amplitude-modulated triangle or sine waves or whatever, only in the rare case of measuring signals it consists of continuous sine waves.


Quote:

Do you really mean to tell me that nobody has come up with anything that will create a filtered version of that data (at a higher sampling rate) that is audibly a clear (to anyone who might hear it) improvement over just playing back those unmodified values with a linear interpolation between the points?

In fact the raw analog signal before filtering looks rather like a brickwall than the schematic triangle waves in the graph (serving for better clearness), otherwise it doesn't change anything on the problematic. Without filtering, you get the amplitude modulation measuring as 3-dB drop-off at 20 kHz. The classic restoration filter (be it digital and/or analog) is geared to sine waves, so you get perfect measuring results in this respect, but lousy transients.

Moreover there are various kinds of oversampling (= digital filter) algorithms which all try to mimic some sort of «believable» waveforms, primarily in that they create a sinc function, meaning a low-pass filter. The problem is, the information about the original waveform lost due to the poor sample rate can't really be regained. So every restoration is just a more or less successful approximation.

 

[bigshot]

Upsampling is most useful when you are processing the sound in some way... For instance if you are bumping up the bitrate, applying some sort of noise reduction filter, and then bumping it back down... This allows you to shape the waveform in higher resolution. For normal playback, it's pretty much a wash.

See ya
Steve

 

[JaZZ]

Quote:

Originally Posted by bigshot Upsampling is most useful when you are processing the sound in some way... For instance if you are bumping up the bitrate, applying some sort of noise reduction filter, and then bumping it back down... This allows you to shape the waveform in higher resolution.

Exactly! And that's pretty much what up-/oversampling does during «normal playback»: processing the signal with an algorithm simulating a low-pass filter to allow a simpler analog low-pass filter. The latter should have an audible effect.

Quote:

For normal playback, it's pretty much a wash.

I think my explanation is a bit more differentiating...

What would you tell to all the developers of sophisticated upsampling algorithms for (kind of) restoring the original waveform, instead of the crude standard «restoration filter»? Have they wasted all their time? After all upsampling makes an audible difference, although I'm undecided which approach I prefer (I haven't heard a filterless DAC so far). Just push the «Audio» button on your 963SA's remote control to experience it!

BTW, I could already hear the difference between oversampling and non-oversampling CD players of the first generation.

 

[Teerawit]

Maybe I'll have to get the G07 AND the G08 to have both types.

 

[bigshot]

I wasn't talking about a simple lowpass filter... I was talking about more complicated dynamic wideband filters that analyze the waveform and make overall changes to it. These are usually VST plugins in high end audio programs. A sophisticated digital crackle reducing filter might benefit from operating at a higher bitrate, so it can operate at higher resolution. It's like blowing a photo up in photoshop to distort it using simulated optical filters like the bubble plugin. The higher resolution makes it easier to hide the technique. An added benefit would be that the tiny "ragged" artifacts around the edges of the effect of the filter would get smoothed out when it's dithered over as it downsamples again. A lowpass filter is basically a lowpass filter, no matter how it's accomplished. That's like raising the brightness in photoshop. A bigger file size would make no difference with that.

For the life of me, I can't detect any difference between upsampled CDs on my Phillips deck and regular playback. It's like the button that turns off the video circuitry to supposedly make the sound cleaner... no real audible effect at all.

See ya
Steve

 

[Leporello]

So, Nyquist and Shannon had it all wrong, then? If upsampling works the way that it is suggested to work in this thread, it very much seems that we are able to create something of nothing.

I was under the impression that the Nyquist theorem works for all signals below half of the sampling frequency. We lose information in the sampling: the frequencies above that. No amount of gimmickry can bring them back. They cannot be present in any way in a 44.1 khz recording. But up to that point the reconstruction filter (mathematically, at least) knows all there is to know about what is happening between the samples.

I also was under the impression that sinc interpolation has been part and parcel of sampling theory since day one (year 1928?). Please correct me if I'm wrong.

Perhaps the NOS dacs do sound different. But it may be the same thing as with vinyl and and tubes: some simply prefer alterations to the original sound.

A few years ago there was an interesting (and lengthy, to be sure) debate on this topic in the Usenet newsgroups. Check it out.



Regards,

L.

 

[JaZZ]

Quote:

Originally Posted by bigshot I wasn't talking about a simple lowpass filter... I was talking about more complicated dynamic wideband filters that analyze the waveform and make overall changes to it. These are usually VST plugins in high end audio programs. A sophisticated digital crackle reducing filter might benefit from operating at a higher bitrate, so it can operate at higher resolution. It's like blowing a photo up in photoshop to distort it using simulated optical filters like the bubble plugin. The higher resolution makes it easier to hide the technique. An added benefit would be that the tiny "ragged" artifacts around the edges of the effect of the filter would get smoothed out when it's dithered over as it downsamples again. A lowpass filter is basically a lowpass filter, no matter how it's accomplished. That's like raising the brightness in photoshop. A bigger file size would make no difference with that.

Steve... you have carefully avoided to take notice of the actual benefit from the digital filter: simpler analog filter and possibility to create better restoration algorithms. Both need the higher sampling rate to work, just like the examples you mentioned.

Quote:

For the life of me, I can't detect any difference between upsampled CDs on my Phillips deck and regular playback. It's like the button that turns off the video circuitry to supposedly make the sound cleaner... no real audible effect at all.

I can hear both effects. They're subtle, but nonetheless clear to my ears.


Quote:

Originally Posted by Leporello So, Nyquist and Shannon had it all wrong, then?

Not necessarily. They just said that it's possible to reproduce frequencies lower than half the sampling frequency, but didn't say in which quality.


Quote:

If upsampling works the way that it is suggested to work in this thread, it very much seems that we are able to create something of nothing.

I don't get your objection -- if it is one.


Quote:

I was under the impression that the Nyquist theorem works for all signals below half of the sampling frequency. We lose information in the sampling: the frequencies above that. No amount of gimmickry can bring them back. They cannot be present in any way in a 44.1 kHz recording. But up to that point the reconstruction filter (mathematically, at least) knows all there is to know about what is happening between the samples.

Mathematics in this case don't care about psychoacoustics. The range above say 8 kHz is severely affected by the classic restoration filter's filter resonance -- the infamous time smearing. In turn the latter «cures» the treble roll-off and removes the amplitude modulation. But as you can imagine this comes at a price in the form of scenario 1.


Quote:

I also was under the impression that sinc interpolation has been part and parcel of sampling theory since day one (year 1928?). Please correct me if I'm wrong.

I'm not sure about that. However, linear interpolation is what's been stated by the audio press then.


Quote:

Perhaps the NOS dacs do sound different. But it may be the same thing as with vinyl and and tubes: some simply prefer alterations to the original sound.

I don't agree on the tubes and only partly on vinyl. My tests have shown that both solid-state and tube (headphone) amps alter the sound to a similar degree. And vinyl records at least in the case of analog and hi-rez masters can sound more accurate than redbook CD.

 

[Glassman]

digital oversampling filters are no way trying to make up any lost data.. it's just mathematics, for various reasons we need the signal at original samplerate in higher samplerate and doing so requires low pass filtering at half the original samplerate.. in other words even if we oversample by a factor of 8x, which is the most common figure for today's DACs, there is no frequency content from 0.5Fs upto 7.5Fs, Fs being the original samplerate.. why 7.5Fs you can ask, well that's because of mirroring effect which takes place every time you change samplerate, if there was no digital lowpass filter used in the process of oversampling, we would end up with original signal mirrored 7 times all over the new passband, filling the whole frequency spectrum of the new oversampled signal.. that's not something we're interested in.. NOS designs however work like that, but because there is no oversampling but also no lowpass filtering at half the samplerate, you always get a mirrored passband around half the samplerate.. in case of ordinary CDs, you have passband from 0-22.05kHz and exactly the same spectrum mirrored from 22.05-44.1kHz.. luckily, today's amplifiers and speakers can usually take these ultrasonic frequencies without going mad.. that wasn't possible somewhere back in 80's..