44.1 kHz vs. 48 kHz - why not use the higher?

[necjamc]

It works in my post, but not the original post. I get a bad request with the OP's, i rewrote the link, and now it works.

[daeron80]

The text version is truncated. Without dither.
Try clicking here.

[Park Seward]

Originally Posted by daeron80

I think that's awfully simplistic. It's choosing to look at audio through one, small, purely theoretical window.

To be real simple: audio sampled at less than 2X the sampling rate will be perfectly reconstructed (without error). Sampling at a higher rate will not add any detail or make the sound more accurate or make it sound "better". It only extends the frequency response of the sampled sound. So if you compare the 20-20k audio, it will sound identical. There is an whole another conversation on the perceived benefits of recording higher frequencies.

That's pretty important and not theoretical.

No circuit is perfect but the math explaining this theory is proven. It's the basis for all sampling.

Perfect reconstruction is limited by filter design and aliasing components. But that does not prove the Nyquist math wrong.

http://en.wikipedia.org/wiki/Samplin...ical_frequency

Interesting that Nyquist's work was based on getting more signals to pass through a telegraph line per unit time in "Certain topics in Telegraph Transmission Theory (1928)" It is limited to twice the bandwidth of the channel. Shannon then picked up the work and developed the Information Theory.

[daeron80]

That's actually only correct for capture and reproduction of mono, steady state tones. If music comprised nothing but unvarying sine waves, Nyquist would be all we'd need to know. As long as you didn't do anything to them in the digital domain.

[O.G. Killa]

Originally Posted by Park Seward

Yes, you add dither when SRC. Also at the output of the summing mixer as Digi says:

"There is a dithering stage in most double precision plug-ins and one final dithering stage at the post master output of the summing mixer. Dither is noise with very specific properties added to the signal in order to de-correlate the noise floor from the original signal so that when length reduction occurs, the resulting waveform does not contain any harmonic distortion or noise floor modulation artifacts."

And:

"More on Dither

We mentioned before that the two most common methods of reducing the number of bits in an audio signal are simple truncation and the use of dither followed by truncation. Truncation means, in our 48-bit result example, that the system would "chop off" the lowest 24 bits, leaving the highest, or loudest 24 bits untouched.

The result of this is distortion in the remaining signal or "quantization error". After the truncation and at higher signal levels, this "quantization error" resembles white noise. As the signal level drops, the noise becomes more correlated (related to the signal) and results in distortion being produced. In the case of a 24- bit system, this distortion is down around the 24th bit — or nearly 144 dB down from full-scale.

By adding dither, the distortion is de-correlated from the signal, which means it is eliminated at the cost of having a slightly increased noise floor. The noise tends to create significantly less havoc with the audio than the unbridled distortion."

http://akmedia.digidesign.com/suppor..._Box_26689.pdf

I'm not trying to be mean here, but do you even read your own posts? You proved MY point in trying to disprove me and you didn't even realize it!?!?!?! Please, you don't fully understand the issue we are talking about so in fairness to other users please stop posting incorrect information.

You DO NOT, nor have you ever, added dither when sample rate converting. you ONLY add dither when changing bit depth (which IS NOT sample rate conversion). Going from 48bit to 24bit, you add dither. Going from 48KHz to 44.1KHz you DO NOT.

Reducing bit depth adds distortion which is why we add dither. Sample rate conversion DOES NOT add distortion and so there is no Dither added. This isn't open for discussion, it's scientific fact.

[O.G. Killa]

Originally Posted by Park Seward

"Another issue is high-quality sample-rate conversion from 96 kHz to 44.1 kHz. This can be done in a totally transparent fashion, but it must be done with care. The possible forms of distortion in a 2:1 conversion, such as from 88.2 to 44.1 kHz, are only those of coloration of the sound and possible aliasing of high-frequency material. The 320:147 ratio implied by the 96 to 44.1 conversion adds an additional problem, which are distortion products equivalent to multiplying the signal by a pseudo-random periodic sequence with a 147-sample period. Figure 4 shows a highly-exaggerated simulation of an improperly implemented 96 to 44.1 kHz downsampling. The original 1kHz sinusoid is accompanied by distortion products at a 300 Hz spacing in both the positive and negative direction."

http://www.jamminpower.com/PDF/New%2...%20Formats.pdf

Again, you aren't understanding what you are reading. If you had only read ONE MORE LINE further than where you stopped... You missed the
"Note: that with sufficient precision, this conversion process can be done without measurable error, but it requires considerably more precision than the 2:1 downsampling"

Which is what I've been saying all along!!!! You don't try to divide 96KHz into 44.1K (what he is talking about when he is talking about the ratio).

You Multiply 96KHz by 147, then divide by 320. When you do that, NO DISTORTION.

Again, please stop trying to mislead people with novice DV forum babble and papers on consumer audio electronics that aren't even focusing on what we are talking about here. You are taking things you don't understand and trying to quote them as fact...

[daeron80]

James Moorer does a much better job of saying what I was trying to say (too simply) when I said "Coarser is coarser, finer is finer." I was picturing in my head a rough approximation of the principle Moorer outlines, but the way I said it is irrelevant under some circumstances. Moorer states it more precisely and relevantly in his article "New Audio Formats" as follows:

"We can talk about the error of the LSB* being a certain number of volts per Hertz. Given this definition, it is then simple arithmetic to note that the quantization error in volts per Hertz is smaller at higher sampling rates, such as 88.2 and 96 kHz, than at 44.1 and 48 kHz. All processing, such as equalization or dynamics, will share this lowering of the error due to quantization."

*Least Significant Bit

I've demonstrated this several times to my own satisfaction by upsampling material from 44.1 to 96 kHz, processing it with EQ or NR whatever, then downsampling it back to 44.1. This is for circumstances where very heavy processing was required, such that ugly artifacts were noticeable when processing at 44.1. I then did both null tests and listening comparison tests. The files almost null, of course. But the differences, while small, were unambiguous: the files processed at 96 kHz exhibited fewer unwanted artifacts, even after downsampling.

[necjamc]

OK, I have a question that maybe you guys can answer then. If a sample rate of 44.1KHz samples the source 44100 times per second, and 96K 96000 times per second, in recreating the sound, why wouldn't the higher sample rate sound better. I am just interested in the why's, because the Nyquist theory and half the sample rate all makes sense, but if your actually sampling more of the source it seems reasonable that the sound would be better.

Please if I'm way off base, correct me and I'll even delete the post if it's misleading. But I'm really interested in that end of the problem.

[spicemix]

If I count eggs by the dozen or I count each egg will I get the same number?

But anyway, all converters run at 6mhz internally and just filter to whatever sample rate you want using the same techniques graphed at infinitewave. It's all SRC, usually with linear phase filters.

And if the bit depth of the conversion is greater than the output, dither is used when truncating to the output format in a good implementation. This eliminates quantization error. Aliasing and imaging distortion and filter ringing are plotted at infinitewave and are also unavoidable tradeoffs, but less of a problem when the output rate is faster. This is also how nonlinear plugins avoid those artifacts.

We need a good FAQ site to throw at people on this it's tiresome.

[O.G. Killa]

Originally Posted by necjamc

OK, I have a question that maybe you guys can answer then. If a sample rate of 44.1KHz samples the source 44100 times per second, and 96K 96000 times per second, in recreating the sound, why wouldn't the higher sample rate sound better. I am just interested in the why's, because the Nyquist theory and half the sample rate all makes sense, but if your actually sampling more of the source it seems reasonable that the sound would be better.

Please if I'm way off base, correct me and I'll even delete the post if it's misleading. But I'm really interested in that end of the problem.

In some ways you can can use the Film analogy. While sampling audio is not like recording film, you can make a comparison between a film camera's speed and a sampling rate. Why shoot at 30fps if you can shoot at 60 or 100 fps? The answer is the eye and brain only take about 18~20 "snapshots" per second. So the eye cannot detect the difference, therefore it doesn't make the faster film speed to have any "perceived" increase in quality.

Now, other parts of the film process might (and can) benefit from having a 60 or 100 fps film speed, and that justifies it. But the "increase in quality" argument doesn't because there is none.

A very smart man and inventor by the name of Dan Lavry has written white papers about this. He is also famous for designing one of the best sounding A/D and D/A converters in the world... at a price tag of $4500 per mono channel!!!!

So, from a theoretical point of view, there is no audible difference between 44.1KHz and 96KHz in and of itself. But, the design of your converters and the type of processing you do to the audio while in the computer can sometimes sound better at a higher sampling rate.

For example, pitch shifting sounds much better when you've recorded at 96KHz compared to 44.1 or 48KHz. That means something like Autotune or Melodyne will sound a little more natural and have a little less "artifacts" at 96KHz. Certain other plugins, like reverbs as described earlier, can be effected by the higher sampling rate.

But remember, the sampling rate is nothing more than a strobe. It's like sitting in a room with a strobe light on. If you increase the speed of the strobe eventually you won't notice the strobe any more and it will appear as though the light is just "on". Fluorescent lightbulbs do this. If you record a fluorescent lightbulb in slow motion you'll notice that it turns off and on 60 times every second. But to us it just looks like it is "on" because our eyes and brain only take 18-20 snapshots per second. Once the Stobe goes faster than around 20 times per second, the light will appear to stay on even though it isn't.

Our hearing can only hear up to around 18~22KHz and as you get older that drops, with most people in their 70's not being able to hear anything about 12~14KHz or lower. So, 22KHz is the upper most extreme any human can hear... so a sampling rate of 44KHz will be sufficient.

But like I said, there are other factors at play, like the filter used at Nyquist and certain types of DSP that can have a noticeable effect on the Audible band even at higher sampling rates.