Thinking about a new hi-rez audio interface?

READ THIS FIRST!

I ran across this paper whilst cruising another forum. The topic was sampling rates. What is optimal? Is 192Khz the Holy Grail?

Read THIS and save your bones for good mics and preamps.

The dude makes some very convincing arguments and has the math to back it up!

TG

The guy lost credibility with me with this statement:

"Sampling audio signals at 192KHz is about 3 times faster than the optimal rate. It compromises the accuracy which ends up as audio distortions."

This is totally false. If it were true, then the lower the frequency of the wave sampled, the higher the distortion level would be.

A D/A must take only two points of data and construct a distortion free full cycle sin wave at as high as 1 cps below nyquist. This is no trivial feat. His argument is that constructing a sin wave with 3 times as many data points (i.e. 192KHz sample rate) will cause more distortions. That is simply not true.

My question to him is “then what happens when you have hundreds of data points for a sin wave?” - which of course is exactly what you have when sampling a low frequency wave like 1KHz or 100 Hz. Which we all know are not distortion problems for A/D/A processes - in fact they are extremely clean because of all those data points representing them.

The “better” sound of higher sample rates is most likely attributed to the A/D and D/A having more data points to represent/reconstruct the waveforms. Nyquist has a great principle (even though it sounds like a cough medicine) but the bottom line is that it requires a chip to have the ability to construct a full cycle sin wave with very, very little information. And I can guarantee you that the reality is that a chip will do a better job with more information provided to it.

But I do agree with YOUR conclusion that some jam up preamps are a much wiser choice for the upgrade dollar.

I see your point. I did some critical recording/listening tests with my EMU 1820M through Event TR8 monitors. I did not go as high as 192k after I made a few tests at 96k. The difference between 48k and 96k was virtually non-existant. Certainly not enough to justify the added load on the PC. I don’t exactly claim to have golden ears though!

I’m sticking with 44.1k or 48k. I believe to get “that” sound, you have to get it BEFORE the D/A process.

TG

Thanks for sharing. It’s a good read. I did tests with my motu828mk2, and to my ears I couldn’t hear much of a difference between 48 & 96. I’m sticking with 48.

Maybe there was not much difference because other things in the chain were masking it? I bet you can hear the difference between 44 or 48 and 96 or 192 pretty easily on really good equipment. I can hear enough of a difference on my cheapo radio shack/RCA speakers to make me want to do everything at the higher rate. At least I think I can… :)

Wow. I just had a look at this guys client list. He probably knows what he’s talking about!

TG

Quote (gtr4him @ Jan. 24 2005,22:05)
Wow. I just had a look at this guys client list. He probably knows what he's talking about!

TG

That and he makes some top notch converters.
Quote (sekim @ Jan. 24 2005,17:15)
The guy lost credibility with me with this statement:

"Sampling audio signals at 192KHz is about 3 times faster than the optimal rate. It compromises the accuracy which ends up as audio distortions."

This is totally false. If it were true, then the lower the frequency of the wave sampled, the higher the distortion level would be.

I believe that the point that the author is making is that sampling at such high frequency leads to inaccuracies brought about by physical limitations of the analog circuitry. He says that slowing down improves accuracy.
Quote (mblackman @ Jan. 25 2005,08:37)
I believe that the point that the author is making is that sampling at such high frequency leads to inaccuracies brought about by physical limitations of the analog circuitry. He says that slowing down improves accuracy.

I read the entire paper again last night and I believe that is EXACTLY the point he was making. Because of the physical limits of analog circuits, higher sampling rates can cause distortion artifacts.

It's all about tradeoffs in design I suppose...........

TG

I thought the point was that nyquist had shown that all information concerning a waveform is captured by a rate twice the rate of the highest frequency.

Quote (TomS @ Jan. 25 2005,10:26)
I thought the point was that nyquist had shown that all information concerning a waveform is captured by a rate twice the rate of the highest frequency.

True, but most people who push the high res stuff ONLY talk about the advantage of SPEED. Due to the physical limitations of the analog circuitry, ACCURACY becomes an issue. Read the paper about two or three times. The author makes some valid points IMO. Not that it really matters as I will probably never have the cash to plop down for a 16+ channel 192K interface. :)

TG

I dunno, TG, the cost on these things keeps coming down.

I only scanned the paper, I will take a closer look at it. However, I am brave enough in my ignorance to ask a question: the physical limitations of analog circuitry you mention, are those absolute barriers, or might there be ways of improving the analog stuff to solve that problem? If the latter, then the nyquist claim becomes more important to the argument.

Like I said, my ignorance makes me brave. :)

One point he makes is that ANY frequincies above Nyquist will cause distortion, and it will be very audible. Most cards don’t have that kind of hard bandwidth filters, so they benefit from pushing up the samplerate to the point that there are no frequencies up there.

Wish I had the math skills to make complete sense of this, but such as they are it’s a very convincing argument. Even so, I’ve seen the bits/sample rate arguments hashed out on other boards in the past and recall another motivator for avoiding the new best sample rate. That being the conversion process to get your end result back down to 16/44 for CD audio is imperfect. Again, the reasoning is well over my head, but the recommendations were to use 44100 or exact multiples and is connected to the “Use Dithering” option that is recommended in the N-Track manual.

I have a soundcard that will do 96K, but I’ll be using the 44.1K setting at least until I have good enough mics/pres/amps/monitors to be able to tell the difference.:slight_smile:

Most cards don't have that kind of hard bandwidth filters, so they benefit from pushing up the samplerate to the point that there are no frequencies up there.
That used to true but isn't with "one-bit A/D" technology, also known as "oversampling", and it's the thing briefly mentioned in the paper where the input sample rate is actually very high internally and then converted to the sample rate presented to the software. This internal conversion includes the nyquist filter, and works WAY better than the analog methods that were used previously.

However, there are probably differences in these filter algorithms, and for any that don't eliminate any signal over 22.05kHz, using a higher sample rate reduces or eliminates the aliasing that would be caused by it. From the paper, I get the implication that these filters are good enough for the job, though.

I'm completely convinced that to reproduce music, 44.1kHz is sufficient, and going higher only helps if there are flaws in the soundcard's design that are offset by this (assuming the analog circuitry is good enough so that it actually works well at the higher rate.)

But I'm not convinced that higher sample rates won't improve the quality of a mix.

Like I mentioned above, the guy ONLY discusses reproduction, and he totally ignores the effects of processing. In another thread, duncanparsons explained to me how most choruses and simple pitch shifters work, using a variable delay line. Due to the fact that "time" in a plugin marches at the sample rate, the only way to do a variable delay line is to skip some samples and double others (depending on whether you're speeding up or slowing down). Without having done the math, I'm pretty darn sure that this process introduces quantization noise, and the level of the quantization noise goes down as the sample rate goes up. Ergo, this is one case where using higher frequencies helps.

Interestingly, in this case, recording at 44k, upsampling to 192, running the plugin, and then downsampling back to 44k would produce better results than staying at 44k (assuming the up/downsampling code is very high quality). I'd always thought that upsampling wouldn't provide any benefits. This is a testable hypothesis -- I'll get back to you on what I find out. (It's also clearly demonstrable with simple graphs.)

the recommendations were to use 44100 or exact multiples and is connected to the "Use Dithering" option
I think this used to be true but isn't these days due to the way downsampling algorithms work, which is to convert the input to the lowest common multiple frequency and then back to the output frequency. The "dropping every other sample" method isn't used any more; it's lower fidelity than the upsampling/downsampling method. I do know at least one famous audio engineers who used to espouse this no longer does.

One thing we PC-based DAW folks have to remember is that a lot of rules of thumb were based on hardware limitations of DAT gear or DSP chip technology, and just don't apply to our case.

BTW, dithering has two purposes:

1) to reduce quantization errors when reducing sample bit-depth. This has nothing to do with sample rate conversion.

2) to mask quantization noise with more pleasant white noise. This might be related to sample rate conversion, but if conversion is done in 32-bit float mode, it's a waste of time to dither because the q-noise is insignificant compared to the q-noise associated with item (1) above.

Ergo, dithering should generally be done only as part of bit depth reduction. There are exceptions, of course.