audiophiles gone crazy (again)

[ErikKaffehr]

Thanx! Interesting!

Best regards
Erik

For all those confused by all the digital audio tech talk concerning quality control through quantization, DAC, high bit data, etc. I offer a more simple but concise explanation by engineer Monty at xiph.org...

https://www.youtube.com/watch?v=cIQ9IXSUzuM

Don't confuse high bit data output with capture input which is the photographic equivalent of claiming more data is seen with 14 bit camera ADC captures processed and interpolated to 16 bit ProPhotoRGB in ACR/LR but viewed on an 8 bit video screen as long as you use a gold plated display cable connection.

How the hell can anyone find the weak link that affects quality in that spaghetti chain of complex inter-connectivity? I've got some special bundled securities I'ld like to unload at a cheap price if you don't ask too many questions.

[Telecaster]

For performing (strictly for fun these days) and recording my main guitar amp rig consists of a tweed Fender Deluxe, a Vox AC15 Twin and a Marshall 1958 (18 watts, 2x10" speakers). All three amps well-maintained, outfitted with taste-tested NOS tubes of the proper vintages and running at 1960s-era voltages via custom regulators. (The Vox & Marshall run at stepped-up ~220v, the former 'cuz it sounds better that way and the latter 'cuz its power transformer doesn't speak American.) My main guitars are a Fender Telecaster (ahem!), a Gretsch Chet Atkins, a Supro Dual-Tone (AKA the "Les Paul slayer") and a Fender Jaguar. All 1950s & '60s stuff…'cuz sometimes the "older is better" clichés are true!

However when it comes to listening to recorded music I can hear a great performance of a great song or longer-form piece via my iPad's internal speakers just about as well as via anything else. I often listen to music this way in fact. I also sometimes listen through my guitar amps a la Leo Fender (he favored a Twin Reverb). In mono, of course. The Deluxe in particular is great for judging mix quality…it'll tell you whether or not the audible bass frequencies are dialed in right. And I've actually done mono mixing via a Magnatone Custom 280 (it has a pair of 5" "tweeters" in addition to the two 12"ers.)



-Dave-

[hjulenissen]

Should we next debate the recent Riedel strategy of selling numerous different type of wine glass, one for each grape variety?

People may purchase an expensive car, a watch or an expensive camera (or camera accessory) for a variety of reasons. I think it makes sense to be aware that we (as complex social beings) make our choices and preferences in a hard-to-predict manner, one that we often do not comprehend ourselves. Marketing people are perhaps "experts" in what makes people purchase something, but even they cannot perfectly predict how a product will do.

Does this justify being a party-pooper, ridiculing the guy who spent $1000 on Riesling wine glasses? I don't think so. Let people be people, and let us have our interests and quirks. Only when that guy finds himself a soap-box, claiming for the world that his $1000 significantly altered the physical stimuli to his nose and tongue, and that anyone questioning his physical explanations should stop doing so does it make sense to enter the debate to me.

-h

[Bart_van_der_Wolf]

How the hell can anyone find the weak link that affects quality in that spaghetti chain of complex inter-connectivity?

Hi Tim,

Thanks for the link.

How, you ask? It's not that hard, if we stick to the basics.

1. We need to sample at a frequency that's high enough to avoid aliasing. IOW, we need to sample at a frequency of more than 2x the highest input signal (Nyquist) frequency.

2. We can store that information in a digital form after quantization. The more bits we use, the lower the quantization noise will be, but we do not need to go to extremes because human hearing can hardly hear the noise below a certain level (and that level also varies other sound levels present at the same time). There are also methods of hiding the noise by filling in some of the non-randomness of the quantization noise with dithering. And there are dynamic compression methods available which allow to use lower precision, fewer bits.

3. Reduce the amount of noise that's added by the equipment that is used in the chain of events leading to output.

4. Make a cost benefit analysis in design, as to where the most benefits can be gotten in the 3 previous steps.

Spending too much of our resources on step 3 without having done some homework on 1 and 2, doesn't make much sense.

Again, these steps are not really different from Digital Signal Processing (DSP) for imaging, the time domain is just exchanged for the spatial domain (usually in 2D).

Cheers,
Bart

[BJL]

Does this justify being a party-pooper, ridiculing the guy who spent $1000 on Riesling wine glasses? I don't think so. Let people be people, and let us have our interests and quirks. Only when that guy finds himself a soap-box, claiming for the world that his $1000 significantly altered the physical stimuli to his nose and tongue ... does it make sense to enter the debate to me.

Agreed!  But getting back to my original post, the proponents of $10,000 ethernet cables were on that soap box (and using it to sell expensive snake-oil), and so I deem them fair game to have it kicked out from under them.

[BJL]

1. We need to sample at a frequency that's high enough to avoid aliasing. IOW, we need to sample at a frequency of more than 2x the highest input signal (Nyquist) frequency.

2. We can store that information in a digital form after quantization. The more bits we use, the lower the quantization noise will be, but we do not need to go to extremes because human hearing can hardly hear the noise below a certain level (and that level also varies other sound levels present at the same time). ...

Isn't that what the CD standard does: sampling at 44.1KHz (EDIT: after low-pass filtering) in order to faithfully handle frequencies up to about 20KHz, and digitizing at 16 bits on the basis of measurements that this goes beyond what the human ear can discriminate.  (Aside: I have read that when Phillips and Sony jointly developed the CD audio standard, Phillips first proposed 12-bit, but Sony provided evidence of an audible advantage in going to 16-bit instead.)

I understand that recording often instead uses an even higher sampling rate, and higher ADC bit depths, so that low pass filters with a gentle roll-off can be applied to the sampled signal before down-samping to the 44.1KHz, 16-bit output format.

[Tim Lookingbill]

Hi Tim,

Thanks for the link.

How, you ask? It's not that hard, if we stick to the basics.

1. We need to sample at a frequency that's high enough to avoid aliasing. IOW, we need to sample at a frequency of more than 2x the highest input signal (Nyquist) frequency.

2. We can store that information in a digital form after quantization. The more bits we use, the lower the quantization noise will be, but we do not need to go to extremes because human hearing can hardly hear the noise below a certain level (and that level also varies other sound levels present at the same time). There are also methods of hiding the noise by filling in some of the non-randomness of the quantization noise with dithering. And there are dynamic compression methods available which allow to use lower precision, fewer bits.

3. Reduce the amount of noise that's added by the equipment that is used in the chain of events leading to output.

4. Make a cost benefit analysis in design, as to where the most benefits can be gotten in the 3 previous steps.

Spending too much of our resources on step 3 without having done some homework on 1 and 2, doesn't make much sense.

Again, these steps are not really different from Digital Signal Processing (DSP) for imaging, the time domain is just exchanged for the spatial domain (usually in 2D).

Cheers,
Bart

Yeah, but how does all that tell you what to listen for on whether it improved the image or sound with so many other variables up & down the pipeline which was the point I was making. Listening involves a complete package that can't be easily dissected to determine whether each has an influence.

We don't judge music according to what a scientific instrument indicates about the integrity of the source and output data and as far as I'm concerned it's too complicated to bother to find out. But I know I don't need to spend a lot of money to get a really good sound which may be reproducing 90% of the true integrity of the source. It's just no one in the industry makes it easy to see spending the extra money for the 10% is worth it.

But thanks for the effort behind attempting to explain according to how you interpreted my statement, Bart.

[Telecaster]

I understand that recording often instead uses an even higher sampling rate, and higher ADC bit depths, so that low pass filters with a gentle roll-off can be applied to the sampled signal before down-samping to the 44.1KHz, 16-bit output format.

Yup. 24 bit/96KHz is typical, though you'll see sampling rates of 48, 88.2 and even 192KHz too. For playback I can't hear the difference between 24/48 and 24/96 (typically FLAC format) but depending on the source I can hear 16 bit vs. 24. (Though sometimes a 24-bit master will have had gentler limiting applied to it than the 16-bit version, which complicates A/B comparisons.)

-Dave-

[Petrus]

If we look at typical good listening systems and rooms and habits, it becomes quite evident that 16 bit sampling for the final delivery format is enough.

First, the ambient noise of the room: rooms with 20 dB SPL ambient noise level is exceptionally good. Most rooms we would consider quiet are in the 30-40 dB SPL range. A good stereo system should be able to put out clean uncompressed 105 dB SPL peaks, this is the K85 monitoring level recommendation (85 dB SPL at -20dBFS). Doing simple math shows that with the 0 dBFS hitting 105 dB SPL output level, the theoretical digital noise level of a 16 bit file is at 105-98=7 dB SPL, which is well below the ambient noise of practically ANY room. Using 24 bit sampling would place most of the extra detail so important to audiophiles not only below the room ambient noise, but also below the threshold of hearing. Besides we can in no circumstances discern detail buried under more than 60 dB of louder signal anyway. The very quietest classical recordings that I have heard had bit over 70 dB of DR between peaks and the recoding venue ambient noise, that is over 20 dB less than what 16 bits can accommodate.

I do classical recoding as a hobby, and both converters* I use are "the best in the world" in a broad sense of the word (difficult to establish the truth in this, if there is one, they are damn good anyway). The signal to noise ratio is around 127 dB or 21 bits worth, which seems to be the practical real world limit at the moment. While I wrote that 16 bit sampling is good enough for delivery, I do record at 24/88.2, not 16/44.1. Why?

Using 24 bit sampling gives a lot of extra headroom at the recording stage. Getting the maximum 16 bit quality means trying the set the levels so that peaks hit almost zero, preferably -0.3 dBFS or so, but NEVER CLIP. Nerve wrecking experience especially in live situations where levels can not be exactly set beforehand. Using 24 bits with a world class ADC means almost 30 dB of safety margin for the same end result. In practice I set the levels about 6 dB lower than where I expect the peaks reach.This brings a lot of safety and peace of mind without quality loss. So the 24 bits is more a safety and convenience thing, not something we can actually hear.

Why higher sample rate? This is more for peace of mind, is a client wants a hi-res file afterwards. There is also one slight quality advantage: the best available sample rate converters** are software based (and free, or practically free), so it just might be so that converting a 24/88.2 or 24/96 file into 16/44.1 with those gives a slightly better end result compared to a in-machine 16/44.1 file.

*) Sound Devices 722 field recorder and Prism Orpheus 8-track ADC/CAD converter
**) Izotope SCR bundled with cheap but excellent Audiofile software, or SoX, which is a free command line driven sample rate converter.

Addendum: I made a test file where part of the signal was original 24/96, part 16/44.1 quality edited in at random intervals (converted to 16/44.1 and back to 24/96). I could not hear which parts were 16/44.1. So I am more than skeptical about the "day and night" difference claimed by some.
 

[hjulenissen]

If we look at typical good listening systems and rooms and habits, it becomes quite evident that 16 bit sampling for the final delivery format is enough.

First, the ambient noise of the room: rooms with 20 dB SPL ambient noise level is exceptionally good. Most rooms we would consider quiet are in the 30-40 dB SPL range. A good stereo system should be able to put out clean uncompressed 105 dB SPL peaks, this is the K85 monitoring level recommendation (85 dB SPL at -20dBFS). Doing simple math shows that with the 0 dBFS hitting 105 dB SPL output level, the theoretical digital noise level of a 16 bit file is at 105-98=7 dB SPL, which is well below the ambient noise of practically ANY room. Using 24 bit sampling would place most of the extra detail so important to audiophiles not only below the room ambient noise, but also below the threshold of hearing. Besides we can in no circumstances discern detail buried under more than 60 dB of louder signal anyway. The very quietest classical recordings that I have heard had bit over 70 dB of DR between peaks and the recoding venue ambient noise, that is over 20 dB less than what 16 bits can accommodate.

I'd note that dB SPL may not be a sufficient metric in this context. Acoustic room noise tends to be highly non-flat (lowpass). So while the broad-band acoustic noise power rating may be e.g. 60dB below signal peaks (either in recording or comfortable listening), acoustic signal-to-noise-power at 1-4kHz (where our hearing is most sensitive) or 12kHz could be significantly higher.

Stated simpler: even if room noise completely overwhelms electronic noise at 250Hz (so as to make it irrelevant), this does not exclude the possibility that the situation is reversed at 4kHz or 12kHz.

Using 24 bit sampling gives a lot of extra headroom at the recording stage. Getting the maximum 16 bit quality means trying the set the levels so that peaks hit almost zero, preferably -0.3 dBFS or so, but NEVER CLIP. Nerve wrecking experience especially in live situations where levels can not be exactly set beforehand. Using 24 bits with a world class ADC means almost 30 dB of safety margin for the same end result. In practice I set the levels about 6 dB lower than where I expect the peaks reach.This brings a lot of safety and peace of mind without quality loss. So the 24 bits is more a safety and convenience thing, not something we can actually hear.

Exactly. Just like having 14 stops of DR in our cameras might not be technically _needed_ most of the time, it allows us to be more relaxed with highlight clipping, meaning that we can consentrate on grabbing the right image instead of constantly working to compensate for camera technology.

Why higher sample rate? This is more for peace of mind, is a client wants a hi-res file afterwards. There is also one slight quality advantage: the best available sample rate converters** are software based (and free, or practically free), so it just might be so that converting a 24/88.2 or 24/96 file into 16/44.1 with those gives a slightly better end result compared to a in-machine 16/44.1 file.

An "ADC" is in many ways a "black box". We may (sensibly) speculate that it will typically contain some clock circuitry, an analog stage converting analog voltages into some discrete representation, a digital stage doing filtering and processing in order to shape the samples into a "standard representation" that makes sense outside of proprietary implementations, and an output stage spitting out data over spdif, firewire, USB or whatnot. My point is that the discretization will typically be fixed at some high-rate, few-bit representation that happens to be the most performance:cost optimized at the time of manufacture. Some internal dsp will likely do the equivalent of resampling in order to give both 44.1kHz and 96kHz.

If this is true, I fail to see how one could cathegorically claim that 96kHz followed by software downsampling is the "better" output option. I agree that after careful testing, one might find that a given set of options are "best" for a particular device, but I doubt that anyone will be able to hear this best-ness in a controlled, relevant listening test for anything but seriously flawed products.

Now, if you carefully design your listening test for maximizing the probability of detection, that is another story. I would try recording certain percusive instruments upclose using wide bandwidth microphones, followed by 2x (octave) or more of pitch-shifting downwards. Eventually, the high-frequency details that are dismissed as irrelevant in discussions like these, will be highly audible.

-h