View Single Post
Old 16th May 2012
  #15
Gear Head
 

Quote:
Originally Posted by UnderTow View Post
Hi Dan (and whomever is passing on his responses),

First, thanks for the extensive response!



I think I didn't make myself clear: I understand the points in your article about the analogue stages but I don't see how that applies to the practical world of modern multi-bit sigma-delta modulator converter designs. My point/question wasn't so much about the much higher rate of the modulator or the lack of any compromises in the digital conversion stages but rather about the fact that the modulator rate is constant for all target sample rates subdivisions of the modulator rate. If the target rate is 44.1 Khz, 88.2 Khz or 176.4 Khz, the modulator is always running at 5.6 Mhz. If the target rate is 48 Khz, 96 Khz or 192 Khz, the modulator rate is always 6.1 Mhz.

The modulator is delivering the exact same signal whether the target rate is 44.1, 88.2 or 176.4 Khz (or higher). In light of this, how do the issues you mention with the analogue electronic components affect the different target sample rates differently? If your point was about a difference between 44.1/88.2/176.4 and 48/96/192 I could understand as they have different modulator rates but I do not understand how this would work between say 44.1 Khz and 88.2 Khz.

I also don't see how software algorithms are affected by the speed of the analogue stages? I don't believe they are. Either the software (and the computing hardware it runs on) runs properly or it breaks don't catastrophically. Unlike analogue, there is no graceful degradation of performance. It works or it doesn't. A computer CPU running at 5 Mhz is no more accurate than one running at 3Ghz. The analogue speeds of the components have no bearing on the digital software accuracy. (Of course there are other compromises on the software/digital side but those do not pertain to the points made in the article being discussed as far as I am aware).



Of course. My comment was specifically about that alleged -3 dB point at 20 Khz (and thus a cumulative loss of 12 dB at 20Khz in the example given) mentioned in the article. That doesn't seem to exist in the real world of modern converters and studio monitor speakers. (At least the ones I am aware of). If there are other compromises being made I would be interested in hearing any more information about them but I don't think it helps the cause against ever higher sampling rates to present arguments that do not apply in the real world of modern converters.



I'm not sure what "very demanding" means technically. To me the only question is: Are the designs achieving what they set out to achieve or not? Based on what I know it seems that they do.

I fully understand what you are saying in theory but it seems that in practise, according to any of the more extensive listening experiments I have taken part in or read about, with good converters people can't actually distinguish between a 44.1 Khz ADC -> DAC loop and a straight wire. In light of that, how serious are these audio problems? Isn't, if not 44.1 Khz, at least 48 Khz more than sufficient for all our conversion needs[1]?

Also, purely on a technical point, the ADC chips I have checked use the same filter slope for 44.1/48 Khz, 88.2/96 Khz and 176.4/192 Khz. [2] I don't mean the same cut-off point in absolute values but the same cut-off point as a fraction of FS. Of course if FS is doubled (or quadrupled) any issues with filtering would be moved even further outside the audible band. I am just trying to figure out what you mean by "very demanding". Surely if there was any overt technical issues with the filter slope/processing load on the systems (like overheating or such), the chip designers would have relaxed the filters slopes at higher rates. Or am I missing something?

Also, again purely on a technical point, it seems the flagship chips I have checked have a transition band on their filters of around 4 to 5 Khz (for 44.1 Khz operation). This means there will be a bit of aliasing of frequencies between 22,05 Khz and whatever the stop-band is (up to 25.6 Khz on the Cirrus Logic CS5381 which was the highest of the chips I checked). Do you see this as a problem or would you consider it a good compromise?

Thanks in advance,

Alistair

[1] With the obvious exceptions of things like recordings for scientific purposes or recordings intended for, for instance, pitch shifting or other such processing for sound design purposes.

[2] With the exception of the Wolfson WM8786 chip which uses a different filter slope for quad rate conversion. (But the same at single and double rate speeds).
Here is Dan's response (Alistair’s words in quotes):
“I'm not sure what "very demanding" means technically. To me the only question is: Are the designs achieving what they set out to achieve or not? Based on what I know it seems that they do.”
)

A very sharp linear phase filter requires very long delay (FIR), and has pre-ringing issue. The shape of the impulse for sharp filter causes higher side lobes amplitude. That only becomes an issue relatively rarely (music dependent) and for people with ability to hear very high frequencies. One can do the digital filtering with IIR, but doing that and keeping linear phase is a problem when the filter slope is steep. There are various schemes to help reduce those filter limitations, and some offer some improvements, but the performance of the stiff filter for 44.1KHz is not up to par with that of 88.2-96KHz filter.

“I fully understand what you are saying in theory but it seems that in practice, according to any of the more extensive listening experiments I have taken part in or read about, with good converters people can't actually distinguish between a 44.1 KHz ADC -> DAC loop and a straight wire. In light of that, how serious are these audio problems? Isn't, if not 44.1 KHz, at least 48 KHz more than sufficient for all our conversion needs[1]?”

For my old ears, 44.1KHz is great. I am trying to accommodate all ears, and there are reports of few people that can actually hear slightly above 20KHz. I do think that 48KHz is pretty good compromise, but 88.2 or 96KHz yields some additional margin. Some audio people don’t agree that 44.1KHz yields a wire-like result.

“Also, purely on a technical point, the ADC chips I have checked use the same filter slope for 44.1/48 KHz, 88.2/96 KHz and 176.4/192 KHz. [2] I don't mean the same cut-off point in absolute values but the same cut-off point as a fraction of FS. Of course if FS is doubled (or quadrupled) any issues with filtering would be moved even further outside the audible band. I am just trying to figure out what you mean by "very demanding". Surely if there was any overt technical issues with the filter slope/processing load on the systems (like overheating or such), the chip designers would have relaxed the filters slopes at higher rates. Or am I missing something?”

It is easy to keep the same filter on an IC and just move the filter cutoff when changing the rate. It is also easier to make a 192KHz IC and add a X2 decimator to come up with 96KHz, and another X2 for 48Khz (or 44.1KHz derived from 176.4KHz). But it is not the optimum. A converter optimized for say 48KHz sampling (24KHz “audio” bandwidth in theory) will not be able to accommodate 96KHz (the noise shaping of 48KHz automatically makes frequencies over 24KHz “full of noise”). A converter that can do 192KHz has a built in compromise- the noise shaper is optimized to accommodate 96KHz bandwidth, not 20-30KHz audio the ear needs.

“Also, again purely on a technical point, it seems the flagship chips I have checked have a transition band on their filters of around 4 to 5 KHz (for 44.1 KHz operation). This means there will be a bit of aliasing of frequencies between 22,05 KHz and whatever the stop-band is (up to 25.6 KHz on the Cirrus Logic CS5381 which was the highest of the chips I checked). Do you see this as a problem or would you consider it a good compromise?”

I think it is a compromise to allow aliasing to the hearing range. This is one of the reasons why flagship converters are not a reference for the optimal quality. Once manufacturers agreed to accommodate 192KHz, they agreed to make compromises. All the 192KHz IC’s also offer 174.6, 96, 48 and 44.1KHz output sample rates; but they are built with a noise shaper for up to 96KHz of signal bandwidth. If they used the same parameters to make the best say 30KHz signal bandwidth (60-70KHz sampling), the results would be better optimized for audio. We don’t need 90KHz signals in audio…


And one more point: I brought the examples of cap charging and settling speeds as 2 examples. These are only 2 examples of many. I don’t have the time and space to go through a detailed analysis of sigma delta. I pointed out that the modulator speed is a decision based on available technology. In theory, faster modulator is very beneficial – it multiplies the effectiveness of noise shaping by a heck of a lot. The modulator rates have been increasing. The old DSD was 64fs and now some IC’s go at 512fs, even higher. So why not 4096fs or more? Because of the drawbacks of speed/accuracy tradeoffs due to the technology limitations (including the 2 analog examples I used and much more).

My statement about the speed/accuracy tradeoff is wide reaching. I am not ready to claim that the principle covers all aspects of life; although usually taking one’s time to do something offers the ability to accomplish a better job. But in electronics, the speed/accuracy tradeoff is solid. Alistair seems to be very pragmatic - interested in end results so he can appreciate what I said- that there is an optimum rate, and that faster does not always result in better accuracy. My main point is that there is an OPTIMAL RATE, not that faster is better. Alistair’s “liking” 44.1KHz actually supports my argument that too fast is ridicules. In this day and age where mp3 is so widely used, I would be pleased to see the CD format hold its ground. And for people that want the highest quality, I want to have some reasonable margin to optimize audio for the most sensitive and critical ears.

Dan