The truth about Nyquist and why 192 kHz does make sense
Vložit
- čas přidán 22. 08. 2024
- I have been attacked quite heavily for claiming that we need higher sampling rates to increase the time resolution up to that of our auditory system. Time to respond. (subtitled in English and Dutch - Nederlands ondertiteld - כתוביות עבריות).
The written English version:thehbproject.co...
Dutch written version: thehbproject.co...
Link to my book: thehbproject.co....
www.hansbeekhuyzen.nl
www.theHBproject.com/en
/ thehansbeekhuyzenchannel
hansbeekhuyzen
google.com/+TheHansBeekhuyzenChannel
Twitter: / hansbeekhuyzen
This matches BBC Engineering's findings when they first introduced digital audio to the broadcast chain (I led a BBC group on digital audio testing). They were also concerned about the effects of cascading multiple A-D and D-A converters, as there was so much analogue equipment between the microphone and transmitter, eg a digital mixing desk, an analogue VT suite, analogue continuity mixer, digital transmission backbone (NICAM), analogue transmitter, etc.
We did a lot of measurements on the early DACs to test for artifacts and came to the same conclusions. Steep filters produce more ringing and have poor impulse responses, whereas a more gentle filter doesn't fully attenuate the signal until a much higher frequency, thus requiring a higher sampling frequency. If the sampling frequency is not high enough, any input frequencies above the sampling rate will reflect back down into audible spectrum causing artifacts.
Thanks Alastair for letting us know. BTW my opinion is based on the research that Stuart and Craven have published over the years. Since it coincides with my auditory observations, I felt confident to publish them in this video form. Extra conformation from you does really help.
Is that why video uses 48khz?
This is why 48khz is used for studio reporting, so that you can have a smoother low pass filter and then the final result can be saved at 44.1
You need to have equipment (including speakers) that can reproduce the the increased frequency range. Most consumer speakers cannot reproduce up to the 20k that 44.1K provides.
@@pilotavery 48 kHz and 44,1 kHz was to avoid digital copies. Eg DAT recorders used 48 kHz which prevents making digital copies of CDs.
Thank you for taking the time to request that all commenters use arguments and remain respectful, and for removing any that didn't. It seems to have resulted in an interesting and useful discussion, a rarity at CZcams.
@averieway Absolutely!
Hans, it's a double-edged sword. Yes, a less steep rolloff would be beneficial, but aliasing occurs in the analog world as well. With such a high sampling rate of 192kHz to rolloff to inaudible levels, there is a greater chance of high frequency noise to exist due to the highest possible frequency content (96kHz). The issue is that this noise is in the nonlinear range of some amplifiers, and nearly all transducers. Thus, the difference between frequencies in this range will be scattered throughout the spectrum, into the audible range. Then, all equipment must be capable of 96kHz playback, which could require items such as supertweeters to reproduce sounds you can't even hear, to prevent audible distortion.
To me there's no obvious solution, as a gentle theoretical rolloff is preferred, but high sampling frequencies causing intermodulation distortion is also undesirable. Your points are fair, but there are definite cons to 192kHz sampling, so no clear answer is available. At least we both agree that we can only hear 20kHz (or maybe slightly above when we were younger), so many people I see arguing for high sampling rates think they can hear 90kHz or something, and tend to seem quite delusional ;)
I agree with you, the higher the sampling rate the more probabilities of getting armonics at our reconstructed signal. What we could use in order to get a better quality is more bit deep in both adc's and dac's.
Jonathan, I think you might be confusing non-linear artifacts (distortion) with phase response problems. Drivers exhibit wonky phase response off axis at frequencies which are too high or low for the type of driver in use. You will hear “non-linear range” while they are referring to the phase response problems. This has nothing to do with harmonic distortion. Distortion is also said to be non-linear, which is where the confusion comes from.
Non linear artifacts happen when we clip a signal. New harmonics not originally in the signal are created. Speakers have specifications for maximum output: peak (clip level of a transient) and RMS (the output level at which constant energy clips). It is the roughly 8-10dB below the speaker’s specified clipping point that we want to avoid, in order to not create harmonics in the reproduced sound which weren’t in the recording.
Speakers and amplifiers do not automatically distort signals at all SPL levels within a certain frequency range (that is unless it is an incredibly bad speaker :). In this way there is no ‘non-linear (frequency) range’. Distortion-induced problems like intermodulation, happen near the top of the speaker or amplifier’s dynamic range.
Another aspect to all this: Music is not flat. Progressively less energy as we move up higher in frequency. This proves true when looking at music on an FFT or RTA. The fundamentals are lower; the harmonics are higher, yet the fundamentals have way more energy. The harmonics have progressively less energy the higher in frequency they go. A 96K high res recording for example, will show on an FFT that 99% of the energy in the music is in the audible band. Comparatively little above 20Khz. This makes it hard to image the ultrasonic range clipping. Everything below 20Khz would have to clip to an absurd degree in order to get the VHF up to the clipping point! You’d have turned your speakers down well before that point.
You refer to noise in the non linear range. You mean noise above 20Khz? For the noise floor to be near maximum output at any frequency, and therefor clip the driver or amplifier; there would have to be something seriously wrong with the amplifier! The kind of problems you are worried about only happen when we are saturating the signal at some gain stage (or at the driver). This does not happen with low level signals. I think some of the extra concern comes form SACD, which was a noisy technology. PCM does not exhibit the same problems in the VHF.
If we say that we shouldn’t have sound reproduction above 20khz because there could be distortion-induced problems, then we would have to say that we should not use speakers at all (because you can have distortion induced problems at any frequency). The difference is us the listeners. And wether a person can hear the difference between a 10,000hz square wave and a 10,000hz sine wave, with an analog oscillator, in a 100% analog environment, with a system that can reproduce VHF….
Never heard above 17kHz in my life, even as a teenager. No music up there anyway
I've done quite a bit of recording and playback at 192khz, and even some at 352khz, and I've never encountered said intermodulation issues. For decades we have heard theoretical concerns about intermodulation distortion, but I've never encountered anyone who actually identified it in a real world scenario (except trying to mix DSD64 on analogue). I have found that converters with higher jitter clocks do not handle 4x and 8x rates as well, but in the age of femto clocks, that is no longer an issue.
I would not doubt that there may be greater ease and audio quality for the listener with less ultrasonic content, but I just haven't heard any significant signal degradation that was greater than the benefit longer filter transition bands. And even so, if the recording was made with transformered mic preamps, or the DAC/Preamp uses transformers, there is generally a rolloff above 50khz and not much reason to even worry.
Thank you for an excellent and very informative video, I enjoy them immensely! I am sorry that you have been treated shabbily, the Internet can bring out the best and the worst in people and I applaud your standing up and demanding to be treated with respect as all humans should be treated with respect and dignity. Bravo!
As my mother-in-law used to say: if you can't' stand the heat, don't come to my kitchen. And about 95% of all comments are flattering, to say the least. Happy Holidays and thanks for your concern.
Thank you, you have clarified some issues for me and answered arguments that there is no validity in sampling above 44.1 kHz.
One correction: In mathematics, "theorem" is an explicitly proven, factual statement. It is a very different thing than a scientific "theory" where there is an explicit question about the universal provability. The Nyquist-Shannon theorem has been perfectly proven mathematically; there is absolutely no question to it's complete accuracy.
There is potential validity to your statements pointing out that it is for perfectly band limited systems, and so aliasing is an issue. But that is a very different thing than implying there is ANY question about the validity of the theorem.
I do not question the Nyquist theorem.
Thanks for great educational videos! Filtering is very expensive especially if you demand high quality sharp response. So it’s ingenious for DSD, MQA et al to use high sampling frequency to alleviate the need for high Q filtering.
Absolutely!
Remark to the band limiting during recording: it is already done by microphone. Almost any of the mics commercially available has lower frequency limit somewhere between 20-50 Hz and upper limit somewhere between 12-20 kHz. For both infra- and ultrasound recordings, specialty microphones must be deployed, which are almost impossible to buy for common people. A friend of mine was doing his diploma work about bioacoustics (bird voice signals, specifically) and even to find, not to buy mics able to go somewhere between 20-80 kHz was really pain in the ass.
Not talking about a need of having covered both "normal" and ultrasound ranges simultaneously.
Although I had Philips earphones able to reproduce ultrasounds up to 26 kHz I never had a mike able to go such high.
It is easily demonstrable: try to do a recording with any of your mikes at any sampling rate and calculate the spectrum at any moment. You will not see any frequencies above approx. 20 kHz with any mike.
(Actually, I can see a problem with the reconstruction of non-harmonic signals, but to find an answer how that is done, I need to do some maths and measurements with signal source - e.g. reconstructinng of square or triangle wave needs theoretically infinite Fourier series. Long story short, not all audio signals are superimposed sine waves.)
Good argument indeed. Thank you for bringing this up, and putting this swing on it.
The fact, that something sticks mathematically doesn't mean that it is technically achievable. The mathematics of the Nyquist theorem obviously doesn't take into account any implementation issues such as how to limit the bandwidth of a signal. In short, one should take into account the practical limitations of the technical implementation and then apply the Nyquist theorem. Of course, in the 80's when the Redbook standard was defined, this would prove too costly, hence corners needed to be cut. Luckily we now live in an era in which there is so much more understanding about digital signal processing and also a standard of hardware that allows us to no longer cut corners.
I wished it were. We still have to cut corners, less than in the 80's, but still... But you can cut corners wisely, dump, cheap and so on.
I believe you are most likely correct in your ideas. What I observe (including my own personal experience) is that people listen to music in increasingly noisy environments, and so the artifacts from lower quality music files and equipment are overwhelmed by the sheer noise of our daily environment. My own daily experience includes listening while walking on a sidewalk with cars a few feet away, on a fairly loud bus, and in an office that has an almost dangerously loud HVAC system (as well as a noisy open office layout). All of these swamp the need for better music files and equipment in my opinion
I agree completely. That is why I use 256 bps AAC on my smartphone that normally is connected to my car stereo. The noise level in cars is so high that details are masked. I only write about stereos in the house. Unfortunately there are many stereos that won't let you hear the difference between MP3 and uncompressed. As can be seen from the comments on my videos. If people hare happy with that, that's fine. But tests I did indicated that when confronted with an affordable but well chosen set, like my Set 3, they are flabbergasted (is that word still used?).
yes, flabbergasted is still appropriate. interesting fact, flabbergast began as a fictional, pseudo-german word meaning, "flutter-ghost"; a reference to bats. so to be flabbergasted is to be bat-bothered (^..^)
I have been using digital audio editing software, and decided to use the Nyquist theorem (with Frequency Analysis) to detemine the highest frequency in my recording, then double that rate. I was trying to reduce the file size of samples so I could fit more on my hardware sampler (musical instrument). The result sometimes was aliasing, which sounded like distortion or smearing of the audible signal, so I think this is correct. Nyquist is the minimum you need, but a higher sampling rate may sound better and reduce aliasing.
I agree to a certain degree. If the really chain is almost optimal, the differences between 44.1 and higher become insignificant.
Great Video, perfectly hitting the spot. But one should consider that our hearing system (ear - nerves - brain) should be able to perfectly reintegrate even signals with distinct pre- or post-echoes, so in most cases these flaws should be inaudible, even for a trained ear, simply because our hearing system also "applies" a lot of filtering, but in psychoacoustic ways. The integrating effect of our hearing system e.g. enables tricky dynamic compression, where waves dont time-align any more, but sound won't change.
8 years ago I got my first ADC that could multitrack jazz trios at 192khz, this same debate was raging back then so I decided to find out for myself. Well let me tell you that my experience was that all things being equal the higher end of cymbals, snares, and female vocalists that I recorded at various rates at the time was preferable at 192khz, and overall music I recorded at that resolution was more pleasing to listen back to. However for practical reasons I can't always choose to use 192khz, but if I could, I would.
I'm just getting into the deeper side of audio, and what you say makes absolute sense. I always figured it was analogous to higher wattage=more headroom in an amplifier, since I wasn't aware of the issue of artifacts. So, higher frequency sampling doesn't remove the problems, just raises them higher than we can hear anyway. It's like filming video: someone could be chatting on their cell phone right outside the frame in a horror movie, but it's outside the frame, so it doesn't matter!
Pretty well explained and argued. I would add that the reason Nyquist works _in_ _theory_ but is difficult to implement _in_ _practice_ is that the Nyquist-Shannon theorems employ a perfect frequency window filter in the _frequency_ _domain_ , whereas practical electronic implementations are limited to imperfect filters in the _time_ domain. This gap between what filtering is required theoretically in the frequency domain and what is possible practically in the time domain, can be bridged by increasing the sampling frequency to a point where the required lower-order time-domain filters become well-enough behaved.
Agreed
In order for me to understand this video I would need another video explaining the jargon galore in this one, like - Time resolution - Impulse Response - FIR Filters - Filter with a Milion Taps - Pole Filter - Jitter - Reconstruction Filter - Band Limited Signal - Filtering producing artifacts - Band limiting reconstruction filter - Aliasing - Proprietary Filtering and Ladder converter.
So I'm afraid I did not understand much what it was about.
I understand that you don't understand. But as I said in the video, you really don't need all this to enjoy music. It was merely to put an end to the 'discussion' about higher standards of music files.
As I said Plexuss, there is no need to know all this just to enjoy the music. Tone it down a bit please. Let's keep this place a nice place.
I totally agree and I will add.
The Nyquist theorem works, and is true, but only to identify sinusoidal signal long enough in time, and constant in frequency and amplitude. In case we approaching the upper limit of the band , and the pulse is short we may not have enough iterations to correctly identify the signal. Example is; the occurrence of one single wave in 19KHz band recorded with a CD standard 44.1KS/s. The real audio consists at most of fading impulses, rather than long in time and continuous in amplitude sinewaves, that is why I consider the CD format to be lossy and a higher sample rate will be beneficial, 48 or 96KS/s. The reconstruction of impulses by DACs is often an approximation. In CD digital format at 5KHz band we get only 4 samples per halfwave, enough to identify the 5KHz wave, but not enough to reconstruct it with accuracy in amplitude and frequency. Increasing the sampling frequency will increase the accuracy and definition, as a proof I can give the equivalent to CD format, stored in 1 bit resolution but iterated at 2.8MHz, the DSD format.
,
I'm a simple man, I see a video with a correct content and I upvote. also I'm an engineer who has worked with DACs/ADCs and I agree, the higher the sample frequency the better you can reproduce/replicate the signal, that is the reason I will not get an oscilloscope with sampling frequency less than 5x of the max signal I want to sample, merely to reduce aliasing.
I recall that in one of the DACs I worked we used a nominal oversampling frequency of 4.8MHz (yes MHz) for most audio, even for something as 16b@8KHz PCM, the only reason for the oversampling was to reduce artifacts in the output signal.
Thanks for moving this discussion forward
I tested this, and you are absolutely right. Although I cannot confirm the theory of course
I want no debate here, we all have our experiences and backgrounds... it’s been awhile since I have looked into these designs, but I seem to recall that using to aggressive a ‘brick wall filter” (digital IIR or FIR filter) when you still had 44.1 kHz sampling can lead to some phase problems after the D/A stage plus the analog reconstruction filter used afterward (which had to have steep slopes). But, usually by late 80s to mid-90s in consumer CD players electronics the combination of an over sampled FIR filter (say 48 k or 96k sampling) combined with a gently sloping analog reconstruction filter was a pretty darned good setup and fidelity! So anyway I like how this gentleman was placing more emphasis on filtering. I like this video.
I will admit one thing Hans and you are perhaps one of the best and most competent reviewers on Hi-Fi currently on CZcams. Your ability to perfectly describe complex phenomena and concepts is unparalleled. Moreover one thing I've always enjoyed about your reviews is they seem to be non-partisan, they don't favor one particular thing or point of view over anything else.
As well I seem to recall hearing that you had a health scare not long ago and I hope you're doing well.
Blush😊
Very accurate and right to the point. Also very clear and easy to understand. Thank you very much
a psychologist would say its a mistake to moderate bad behaviour. The rude man will most likely help you see the answer you seek is in your question.
Technogeek cerca 1984 in the midst of a discussion about the merits of digital audio vs analog audio:
"Anything digital is superior to anything analog."
What a brilliant engineer and mental giant! (Not!)
This video makes a solid point about implementation. I've been saying similar for 35 years.
I am used to hear a lot of non-scientific praises of high sampling rates in consumer audio, but this video actually touched on a subject I never heard addressed, and it is a real thing. Thank you! I am not sure if it's a big problem though. With 48 kHz sampling rate (this is what I sue in my home studio) we have quite a lot of leeway to use less steep anti-aliasing filters. I guess that even steep elliptic filters could be doing hardly any harm to the signal in the audible range at that point. Also: there are other ways to filter digital signals, for example Fast Fourier Transforms (but that probably introduces pre-ringing as you mentioned, same with linear phase filters).
Dig into it and listen. Really!
I didn't know about the 'reverse echo' from digital wave filtering! However I have noticed a great difference with sound quality while working in the 96 khz 24 bit environment of the MPC Live...
I'd like to address a problem that I myself have while doing audio signal processing, and it is that sometimes it needs to be louder in either amplitude or sampling rate for computing purposes, it's not the same to have a sinewave at 16 bit 44.1khz than it is to have a 24 bit 96khz sinewave compressed down to the dynamic range of 16 bit 44.1khz (with a compressor, not reducing the sampling rate) the harmonics obtained out of the process are impossible to replicate from the bitrate of 16 to 8 after normalizing back to 44.1... do I make sense? Think about it like those CSI scenes where they try to resample the face of a man from a very low resolution image back into the crisp and clear high res image. The responsiveness of effects is way bigger while working on a higher bitrate, and higher resolution environment. Other than that, the song is going to sound pretty much the same once it gets it's final mix and gets mastered for distribution. So for production purposes, 24 bit is greater, provides with more virtual headroom, and 96000 khz should be enough, unless the intention is to destroy the sound to extract the very deep harmonics of it, in which case, having more fabric to cut from is always better. :D
So I don't know about going all the way to the 192 khz, or the 32 bit-rate but I assume there could be a benefit in digital audio processing, if the equipment was there to produce the amplitude of such signals at all!
yes, for me as a rookie in this topic, sounds same like i am thinking.
Hi, thank you for the nice explanation.
I hope I have understood (also from the other comments):
1) the recording and mastering is (hopefully) done at high freq (with low artifacts).
2) When creating the CD, the signal is bandwidth limited digitally (introducing artifacts)
3) then downsampled to redbook file.
4) The DAC at the consumer side takes the file and upsample before converting, to reduce reconstruction filter artifacts.
Upsampling in the DAC is key for a consumer...
The improvement in keeping higher sample rate end2end is related to point 2): avoid the digital bandwidth limit filter before downsampling. If only we could have perfect filters, redbook would be enough.
Did I understand correctly?
I am afraid you do really need to watch my playlist on MQA: czcams.com/video/r_wxRGiBoJg/video.html
I think you explained this really well. Very good video.
Hmm... I though CD players solved this already in the 1980s. I mean by having a couple extra bits (say 18 bits) in the D/A-converter and using these to create new interpolated values between the recorded 44KHz 16-bit samples. Hence moving the need for filter cut off as well as filter steepness up (two octaves in the case of 16->18 bits).
The bit depth works in the amplitude domain and the reconstruction filter in the frequency domain. So I'm afraid you are incorrect.
@@TheHansBeekhuyzenChannel Of course, but the time and amplitude domains can be closely connected regarding physical implementation.
The extra bits enables a few extra levels needed to define some interploated values between the recorded samples. Synthetic sample values with shorter duration and thus at a higher frequency. This moves the digital artifcacts up in frequency when fed to a D/A-ladder (with these extra bits) and so relaxes the need for a steep and phase distorting filter. That's the connection between amplitude resolution and the frequency domain here.
A guess you get the point? This principle is what they called "oversampling [filter]" in the late 80s and early 90s.
@@herrbonk3635 Smooth ringing is better than rough ringing... :) It still rings though. The problem arises at the A->D conversion level, extra information is added that cannot be corrected in the D->A process. It's not a lack of bits, but the presence of added HF ringing that is the issue.
@@realworldaudio Ringing? I thought it was the phase distortion itself that was the main problem with steep filters close to (and therefore into) the actual audio band. If you move the filter to 44.1x4 or 44.1x8 (as I described), you would not have any filter effects whatsoever on the audio part of the signal. That's simply why "oversampling" (misleading term) works.
But from your nick name, I suspect you perhaps is the type of guy that don't even want a treble control in your HiFi (a low pass filter as well). These dudes (often called "gold ears" here) somehow seem to belive recorded acoustic music itself is clean from all kinds of artifacts... or that all artifacts always sound really bad, and must go, despite the fact that we have listened to them as long as we have had musical instruments and concerts :D
Wonderful explanation! Many thanks for sharing your knowledge with us. I will order your book forthwith!
Extremely wide bandwith is useful to 'shunt' artifacts out of the 20Hz-20kHz band. This bandwith gives us 'elbow room' working with signals and results in 'cleaner' sound. Some authors posit that after this is done the 'beyond 20kHz' area can be chopped away, retaining the clean part for music reproduction. That is, a properly crafted audio PCM file can be 16/44 or 24/48 after all work has finalized and we have a final master.
Would you be so kind so as to offer us your view on this?
Some people argue that all the 'spikes' and 'brush' in the ultrasonic range may hurt the sonic result if left there for the DAC to process. It wouldn't add anything and it might detract from quality. Once the 'main tranche' is cleaned up we may be able to attain a lower noise floor where it matters by filtering away the 'excess'.
Is this argument 'sound'?
An argument for 24/96 files is the broadening of dynamic range. (The non technical but commercial problem lies in 'the loudness wars' and severely compressed masters and 'remasters')
Modern technology permits widespread use of 24/96 which I hope will become a 'standard' of sorts, at least in the streaming services.
Let's hope quality 'uncompressed' masters find a place in the market. Is LP vinyl dynamic range too much to ask for in contemporary times?
Better sound using higher sampling rates have nothing to do with hearing higher frequencies. No one will argue that 44.1khz is too low. You will get all the frequencies up to 20kHz but unfortunately the low pass filter of 44.1khz can and does cause problems. However a well designed 44.1khz low pass filter will sound better than a cheap poorly designed 96khz filter.. 192khz low pass filter will sound better than a lower one but only if it's done properly. On a cheap $100 U.S. U.S.B interface it is better to record at 48hz than say 96khz. What you say is 101% true but only if done properly.
You make a lot of sense. You are the first person I have heard to describe the real reason why high sampling rates are needed. I was very impressed. This area, although I understood it, was always a little mmmmm.....foggy. thank you for clarification.
It was very pleasing to hear this subject explained so clearly. I decided to digitize my music collection, and yes, to compress much of it for my retirement pleasure. I had a pretty good system, and I just wanted to find the rate beyond which I could not tell the difference. I ultimately settled on 192 kHz. There was nothing technical about the decision, just a desire to get files as small as possible without audible changes. Now I know what I was hearing at lower sample rates. Excellent video!
A perfectly brick wall filtered signal has tons of Gibbs phenomenon, ringing at the cut off frequency, that's the "pre and post echo" as you called it. If you don't have it, then you can't reconstruct the signal and get the peaks and phase right. This isn't really about "digital" it's the nature of frequency. This ISN'T a limitation of technology, it's a limitation of MATHEMATICS. Yes, you're right that you can get rid of the Gibbs ringing if you have use gentle filters instead of brick wall ones, and in order to use a gentle filter, you'll need a much higher sampling rate so you don't get aliasing. But you know, the signal without that visible ringing DOESN'T have less energy at those frequencies, it's just that they're NOT AS OBVIOUS IN THE GRAPH, because they're being compressed into shorter spaces by the presence of higher frequencies.
I don't know what filters are being used in equipment these days. If you use (really sharp) linear phase filters to do the sampling and the reconstruction you won't get any phase artifacts, but one problem with that is that you will get some delay in the signal - that sort of filtering can't even be done in analog. In order to get perfect 44khz sampling you have to actually sample higher than that, filter in the digital domain with some buffering and delay and subsample in the digital domain. I presume that's what's being done.
My own belief is that sampling at these lower frequencies is good enough. Even when I was a teenager I couldn't hear 20khz.
Not only does the Nyquist Theorem require a band-limited signal that you discussed, it also is only for continuous tones. It can be for a multitude of tones of different frequencies and amplitudes, but they must be unchanging over a number of samples. Thus the theorem is not mathematically sound with respect to most music.
Yes, but we are talking music here. Ever heard music that is only a few samples long?
You always make a lot of sense to me Hans, due to your logical explanations (and your English is good too). I know it's easier said than done, but try not to let the attacks bother you too much.
They don't bother me, they come with the job. But thanks for your concern and your compliments.
It would be great if you explain why more than 16bit resolution is needed as well.
I’m not talking about noise. It’s the damage the reconstruction filter causes to the time information.
Sure if the test is set up properly. But that’s not easy
Fully agree!
I know from experience that switching from 44.1 to 88.2 and 176.4 (or 48-96-192), opens up and sweetens the sound, also giving it much more room. I suspect that’s not the high frequency content (can’t hear that anyway), but the accuracy of the timing, that better reflects reflections in the room, and less filtering artifacts. All using proper gear (DPA mics, Jensen/John Hardy mic pres, UA 2192 ADDA).
So I’m interested in the timing issues, and especially in attack. Some instruments (think harpsichord, guitar), have a very sharp attack, which could occur right in between two samples. What happens then? Are you going to miss out on the attack, or will it be time shifted?
Steep filters are always a big problem. Microphones will solve the filtering issue when recording and playing back at high sample rates, but there will be plenty over 20kHz content with good mikes. And when you have to put that on a CD, during downsampling filtering is needed in the end…
I actually chose my DAW based on the quality of the downsampling, which had much to do with timing and filtering. I ended up with a certain German product, which sounded much warmer and retained more than any other product I tested the rhythmic pulse of the music. And sure conversion takes its time, it’s terribly slow.
👍🏼
@MF Nickster Thanks for the hint, I’ll study it.
I think I would want to put an easy way to understand Nyquist theorem. It simply says that the system needs two samples to know that there is a sine wave one on the positive side and one on the negative side. Now I could fit any continuous sample on those two samples. But given that you don't allow any of the higher frequency of those two delta samples to appear, They are pure sine waves when passed through that LPF. Lower frequency components have more samples reducing the source of error and increasing the ease of interpolating the points to a nice clean sine wave by filtering.
I have covered this in several videos, like this one: czcams.com/video/geaoEt-9V-w/video.html
The arguement seems to have merit. Using a higher sample rate ADC so that a more gradual filter can be used makes sense; however storing the audio as 192kHz doesn't. FFT, remove components above 44.1kHz, inverse FFT, and use the standard format. What's the point in storing the data that you don't want and won't reproduce. Even the DAVE claims the standard frequency response of 20Hz to 20kHz
When you downsample to 44.1 kHz you have to use a brick wall filter again and if you want to upsample again it's a reconstruction filter. The clever thing about MQA is that it uses about the same bandwidth as LPCM 44.1 kHz but does keep the higher sampling rate while throwing in compatibility for 44.1 kHz equipment.
In mathematics, a theorem is a statement that has been proved, or can be proved.
Philips and Sony already know of the aliasing problem when designing CD RedBook, that's why they bump up the frequency to 22050Hz, and the sampling rate to 44100KHz. They already take into account the filtering errors. The 44.1 kHz sampling frequency allows for a 2.05 kHz transition band, that is more than enough. I can accept 48KHz that will allow a bit more of headroom for the cutoff filter, but above that leave it for mastering and mixing, but will not improve the sonic performance of the final record.
I know what a theorem is. The reason for choosing 44.1 kHz was primarily to be able to use a U-Matic VRC to store the digital audio signal in a video signal using the Sony PCM-1600 encore (In the US is was even 44.056 kHz because of NTSC). The difference between 44.1 and 48 kHz is negligible. The real reason is that according to Nyquist and Shannon the the attenuation at 0.5fs should be 96 dB (for 16 bit). That means that within 22.05 - 20k = 2.05 kHz 96 dB attenuation is needed, meaning a filter that does 96 dB/prime (single white note on the piano), about 8 times steeper than 96 dB/oct. Just try to find someone that can make such a filter that sounds right!!
Great video! This concept makes perfect sense to me. The bain of speaker design is the Crossover. So why would it be any different with source files? I have vast experience in Car Audio, where we attenuate crossovers readily. From my experience the best sq comes from a more relaxed filter roll off.
Another thing, blending a subwoofer with mids is a real pain in the ass! Why?? Crossovers! Cheers
With crossovers, you can actually do them in the digital domain and have seperate DAC+amplifier for the different bands (bass, mid, treble, or even more of them).
Using a somewhat powerful DSP should give you access to a lot of tricks & math to deal with gain flattening over frequency, phase e.t.c.
For storage, I personally don't see any need for more than 48 kHz 16 bit-ish.
If you want to downsample with digital filtering at recording (using a good filter-algoritms) then go for it.
Same for playback, if you want to upsample and interpolate, by all means. At least you can keep the filters digital and rather complex if so desired.
For me it seems like a discussion about filter type/algoritm vs the other ones.
Thank-you for your videos. Very informative and mentally challenging on the theoretical level.
Danke Hans!!!
I concur with you fully on Nyquist and the validity of higher Sampling Rates and Higher Bit Resolution.
I use 192KHz 32 Bit Float on all my critical recordings and store all my music in that format.
I use software Applications like Audacity for Listening and Export to Established Standards.
I am Not an Audiophile.
I am a Nashville Recording Engineer and Pro Audio Consultant and Studio and Facilities Designer since 1969.
If you check out the etymology of the word audiophile you must conclude you are one😄
I may be a “Closet” Audiophile.
But I am not prepared to come out.
My taste in Music Produced by others tends to follow those classic projects of the 1970-2000 Period except for the Classical Works.
Modern Composers like Aaron Copeland I like very much.
I have always been focused on the end product and I have left the Art to The Artists.
Thorough knowledge of the Technology is essential to my Life’s Work.
I have received the AES Lifetime Achievement Award and I am one of The Architects of the Nashville Sound.
My works with John Hartford and his Dobrolic Plectural Society Band including Norman Blake, Vassar Clements, Tut Taylor, Randy Scruggs, Sam Bush and others are my Favorites including “Aereo-Plain” the First NewGrass Album along with “The Ballad of Calico” by Kenny Rogers and The First Edition.
My work is all Analog using Flickinger and MCI Consoles, MCI, 3M and Scully Tape Machines with Dolby A Noise Reduction and JBL and Audicon Monitors.
I have no back ground in the science but I'm adj and the logic of higher sample is solid more and resolution is what matters not the science
The question is then why not sample at even higher rates. That will lay a burden on all the electronics. It's like carrying all your money all the time in stead of putting it on the bank. That way you can have access to your money all the time? Makes sense, doesn't it. Well apart from practicalities like where to store it, you are risking losing it or being robbed. So it makes sense to carry only the money you foresee you will need. It's the same with sampling. You should only sample the information you need.
This is fine, sort of, but you’ve completely forgotten or ignored in the “recording”
The recording side is not something the consumer has influence on.
There are also those who by using signal generator, two oscilloscopes and spectrometers demonstrate that analog signal is reproduced without loss after being passed through A/D and then D/A. Input is exact as output as observed by one's own eyes, oscilloscopes and spectrometers, so long Nyguist theorem is upheld. Truth be told, people who do such futile demonstrations are people who are influenced by Nyguist theorem and not by your ears. Maybe you have in your possession equipment capable of demonstrating supremacy of your ears, I would not know that, but would love to see some demonstration. Also, let me know if you would like to see demonstration of Nyquist theorem.
Very good approuch indeed, i have to agree with you on this issue, maybe the secret of better sound is not on sheer resolution but outside the range filtering.
I always hear the same brittleness from PCM digital audio no matter what sample rate. The only thing where I hear a significant difference is in albums recoded in DSD DAWs. In the studio I flip the live analog return from the console to the digital return of the DAW through the Digidesign 192 converters and there is a world of difference.
See my video coming Friday.
Hi Hans what this this mean for music reproduction and the file properties needed? If I understand this video correctly, 192 kHz is the sampling frequency for use in the studio or wherever the recording is done to reduce artifacting by aliasing. After mastering by the audio engineer the artifacts is removed.
Is there any audible difference in buying a 44.1kHz/16bit and beyond that? Most digital stores are now beyond CD quality and have similar prices.
I am new to audio and would like to enjoy the music, but would like to understand before spending much money.
There's definitely a difference between the sounds in a DAW. Render and null and discover the difference lies in higher frequency content in higher sampling rates being maintained . Indisputable since its a null of the same material. I only wonder where it matters more since everyone simply starts talking about the subject without clarifying ; are you recording, mixing or listening. In mixing it's quite clear.
You might be right, but do you know what caused the difference. Might it be the reconstruction filter in your DAC's have an easier job? 2L and Channel Classics both offer free downloads on several sampling frequencies and these will not differ from what was in their DAW. When I play them on very high quality gear, the difference is so small that it might just be the result of the downsampling.
When you make copy, the way you judge how good the copy is to compare with original and not the two copies done with different quality or with your ears.
Thank You.
I’d like to know if there are differences between natural instrumentation and completely electronically-produced music in relation to this? Are the benefits just as justified if the instruments are digital to begin with? Is this just for analog instruments (which I know can include the human vocals)? I want to argue for higher bits and sample rates from electronic music but some artists don’t seem to believe in it. It might be linked to the human frequency response thing. Do you still get time smearing in electronic music? Thanks.
Digital instruments vary in quality depending on their working principle. Old analog synths use analog oscillators and filters. They often use the same oscillators for the entire spectrum they cover, making them sound very dull. Although there are those that love the old analog synth's. Nowadays often samplers are used that use pieces of recorded instruments. So there is no synthesis. The quality of those instruments depend on the sampling rate and bit depth, the quality of DSP's they might use and the quality of samples the user loads. If musicians choose a given electronic instrument, they specifically want to use the sound it gives. Which might include time smearing.
Thank you Sir Hans!
I've heard hi-res files played on some sophisticated gear and yes, I found the sound richer.
Enjoy!
Hello, thanks for the video, It was really useful, I have a question, why is important the reconstruction filter in the dac if the frequencies higher than 20 khz are not audible?
The reconstruction filter is an essential part of digitising. It prevents aliasing, misinterpretation of the digital information.
Bedankt meneer 👋
I like hans if you come to Australia i would love you to visit. I could listen all day.
If I ever visit down under, I'll drop you a note.
Aren't we arguing about only a part of the problem? When we record, we don't typically record with a single mic what will eventually be the entirety of an output channel. Has there been a discussion of the benefits of oversampling in a world where multiple digital streams will be mathematically combined (mixing) to produce the output? Even if your final output is Redbook, working at higher sample rates and bit depths intuitively seems to give the algorithms better data to work with before the final mixdown and downsample.
Then again, intuition isn't always my friend when it comes to higher math.
To omit the discussion: just listen.
That's very Zen, my friend. I learned a lot from your video, thanks!
I think a lot of the animosity against higher resolution audio stems from poor performing built in audio on pc motherboards and their high samplerates being used mainly for marketing since the actual analog output sucked. Once again that's what I think.
Or perhaps it's because bit depth controls nothing but noise floor and sampling rate needn't be any higher than 44100 or perhaps 48000 to capture the full audible range with ample room for low pass to take place without detectable aliasing. Hi-Res audio in the listening phase is a waste of space.
Totally agree with these principles Hans. I must confess I didn't believe in Hi-Res audio since the 90's until few years ago.
The biggest advantage of Hi-Res is that the D/A process becomes significantly less critical in the system, in other words, any decent DAC will do a great job when fed with Hi-Res stream and the bottlenecks will be somewhere else in your HIFI system. One downside is the bigger size of the file, but this is becoming less critical with time, and new Codec developments are also helping (e.g. MQA). Another downside is that the "player-making-industry", including the DAC chip makers, have a problem.....how will they differentiate their products if they all perform the same when loaded with Hi-Res?
Pau Rey No argument there. But Hi-Res done right. Not upconverted from a compressed 16/44.1 file which has happened. Or taken from an over Equalized compressed 24/96 file. Hi-res music should be a straight copy from the 32/384 (yes, they exsist) 24/192, 24/176.4, 24/96, 24/88.2 or the 24/48 files. NO COMPRESSION, EQ, OR ANY OTHER AUDIO ALTERATIONS. Expect for edits and I suppose fade outs.
Getting copies of the original master file is what we need to strive for.
I been working in a mixing/mastering studio for 16 years now. And I can tell you that there is a big gulf between what you hear in the studio and what the customer ends up with.
For example: The CD you have of say, release "X" sounds nothing like the 16/44.1 CD production file of "X." They are bit for bit the same but they sound nothing alike. And we don't know why. I have heard 16/48 masters mixed back in the early 80's and I can tell you they sound incredible. Way superior to their DCC or M.F.S.L. counterparts - Not even close!
Don't judge 16/44.1 by commercials compact disks.
Fascinating video and enlightening comments. While I truly am a fan of seemingly esoteric topics (if for no other reason than intellectual curiosity) I think there is a salient point missing here. The number of people for whom a higher sampling rate (and/or bit depth) matters is fairly small - perhaps one in a million. Whether we like it or not, the vast majority of people consume music on either their smartphones with cheap headsets, car stereos, their laptops or, now, digital assistants like Amazon Alexa. Furthermore the music is listened to on busy streets, crowded public transit, noisy offices or other equally noise-ridden areas. For these people, who I would argue make up 95% of all music listened to, the distinctions being discussed here are totally lost. That's not to say we shouldn't strive to decide on the best standards but we just need to be realistic that for most people it just doesn't really matter.
I don’t know where you got your figures but I can assure you they’re wrong. My country has 17 million inhabitants and I surely have more than 17 people in my social circles that love higher quality audio.
The Hans Beekhuyzen Channel I think you are taking me too literally. I was thinking this discussion was broader - discussing what the best recording format and storage is for digital audio. In that light, the large commercial record companies and distribution services like iTunes will make decisions based on what the masses want. Whether the number of audiophiles is one in a million, one in a thousand or one in a hundred, it is certainly less than 5% and I suspect less than 1%. At these numbers, the dominant companies are not likely to be driven to change no matter how compelling the technical reasons. That was the only point I was making.
Remember Betamax? It was by far the better video tape format but died because the general public didn’t care and preferred VHS.
I really enjoy your channel. Just want to point out that a it makes little sense to say that a theorem is only a theorem. The beauty of math is that something can be known to be true with certainty without being empirically proven. The Pythagorean Theorem is also "just a theorem", and thanks to it you will be able to know the exact length of the hypotenuse, if you know the length of the two other sides in a triangle. The way you express that it is "only a theorem" seems like a confusion with the term, from the empirical sciences, hypotheses.
In the context of this video the theorem is just a theorem since it can't be implemented in real life. Especially for non scholars I had to be clear what the term means.
@@TheHansBeekhuyzenChannel Very well. I see you point. However, as a former math teacher it jarred on the ear.
I appreciate that.
Stefan, I am a mathematician. It can be resolved by reading what the Nyquist theorem says, PRECISELY. It does not say that a digital approximation to the perfect sampling in the theorem will give the original waveform back. As simple as that.
@@SveinOlavGlesaaenNyberg Sure. But how is that relevant? Did you watch the video?
Wonderful video sir. I have watch it a few times and I believe I understand what you're getting at- that we need better/more gradual filters above the cutoff. What I don't quite understand is why the filters need to extend so far? All the way to 96khz??? If there is not even any relevant information much higher than 20khz, where are the artifacts coming into play? Put another way- if the filter starts well above the cutoff, does the filter still effect frequencies BELOW the cutoff?
Yet another question I have is.... are there not downfalls to recording/playing back information well above the auditory realm? If I remember correctly Lavry suggests playing back ultra high frequencies is more likely to cause distortion and other issues, rather than improve audio quality. I realize this refers back to the old arguments of higher samplerates... but my point is, isn't it a trade off??? Which is better- to have fewer artifacts from filtering? Or fewer artifacts from introducing ultrahigh frequencies that are not even needed?
The audibility of filters is complex but the steepness of the filter plays a role and the cut off frequency plays a roll in where in the frequency band the most audible artifacts occur. So choosing a higher cut-off frequency and a less steep filter is beneficial for the sound in the audio band.
Thank you for the well thought out response and for the videos. Well done! I'm still skeptical as to the audibility of these filters, especially well above the audio band, and whether or not the benefit/cost ratio is even close to being worth it.. But I would love hear some more descriptive words as to what the artifacts sound like and what to listen for. That said, I'm all for better audio if and whenever possible.
See it like this: the speed limit in my country is 130 km/h. A car that just does 130 km/h needs a lot of time to get at that speed. A car that can do 200 km/h not only gets to 130 km/h faster, it also is quieter, has less vibrations and so on.
First: yes, filters at higher frequencies do have problems that fold back into the audio band. That is the essence of my video. Choosing a higher filter frequency makes sure the artefacts are still above the audio band.
Second: you should not record more than necessary. For us humans the extreme highest frequency young people can hear is 20 kHz. As said, many mikes don't even get that high. So we're effectively recording thin air when recording at 192 kHz. All true. But it does reduce the anti-aliasing and reconstruction filter problems to a very low level.
Higher sample means more room to have more audio bands.
I think part of the discussion that needs to really be explained is the effect of band limiting in both the digital and analogue domain, any band limiting anywhere will causing ringing artifacts, any cable and equipment you use will do this as well. This is why I find double blind tests of audio not truly scientific to disprove high-resolution. If you band limit a cable to 10k due to capacitance that will have an audible effect and therefore you really are not isolating all variables scientifically. The provenance of the signal is no longer intact.
When I hear Pro-Audio people, or people like Monty, tell me they cannot hear the difference, I would want to ask them how many looms of cabe and what kind of monitors are you using? what are their crossover freqs? Any of those things can introduce tons of ringing artifacts that will make high res compromised, or really any signal.
To add to your point Hans, I remember dCS had a white paper on their website some time back, that explained what extreme band limiting in the digital domain did, and they showed it caused phase and amplitude shift, they had snapshots of it on oscilloscopes. This is known in physics as the Gibbs Phenomenon. It is never explained by the apologists of just sampling at nyquist. Any filtering you do will cause a ripple effect in the audio much like throwing a rock in a pond. Humans can hear the harmonics of fundamental frequencies way above 20k and the phase and ringing artifacts smear in the time domain. All of these things are important. If high order harmonics were not important then, for example, the type of piano or guitar you use and how it resonates would be inconsequential because they are all generating the same fundamental frequencies, but as any musician will tell you, it does matter. Why? due to the distribution of how those frequencies sit, and all of this information happens in the upper end of the frequency range.
Personally I have heard the effects of high sample rate and DSD, but it does require a controlled set up, actually I found that tube gear demonstrates it the most due to it's high slew rate, and when I managed a high end store, we had some tube preamps that could go to 100k and high sample rate was very audible on those especially DSD. I also had a CD player that had a band limit of 80k and could select the CD layer at 16/44.1k and the SACD layer, I was even able to demonstrate to audiophiles who didn't believe in DSD that it made a difference it was very audible on a non band limited player.
In short I think the people who maintain there is no difference either have never heard live music or lack the equipment and discernment to hear the difference, because the changes are very audible even to the un-initiated on the right system.
Simply put Nyquist is sufficient to prevent aliasing, but not sufficient to portray real life, and we are after it sounding like real life, not to just no longer have aliasing artifacts.
To use another analogy 24fps will allow the human eye to perceive motion in still pictures, but it doesn't look as if we are actually seeing through a window in real life. The only thing that can approach that is a high resolution screen at a high frame rate, and that is what we want with audio.
You are correct, of course. But the main reason I wanted to publish this video is to silence people like Ethan Winer that posted a remark that "The presenter of this video doesn't know what he is talking about." And consequently misused the Nyquist theorem to prove me wrong. Everyone with ears and a proper stereo can hear that a good 24/192 recording sounds better than a redbook spec file. I don't want to become a tech-geek channel, I want people that are interested in perhaps buying something better than a bluetooth speaker help find their way. Again, that doesn't mean I don't agree with you, but this is not the place.
I do appreciate your outlook a great deal and it is a very good thing to spread the word of high quality audio to the masses, in fact that was one of my great frustrations when I worked in High-End audio that the industry didn't do that enough and try to reach the next generation of people my age. It really is when all is said and done about getting lost in the music and High-Res is a great vehicle to to that even more, and that is something any human being an get behind. I just simply wanted to echo your points as well and perhaps add some to the scientific side of the discussion. I do think the Gibbs Phenomenon in general is a smoking gun for most of what people deem to be "voodoo" in the industry including High resolution, and I think it gives a scientific reason and phenomenological answer to the naysayers for why any of these things matter.
If people are calling you unscientific then I think they need to review the science on all sides a bit harder. I think the problem comes in when people who aren't trained in the more modern side of engineering come in a say otherwise, when in fact the ability to even measure these things as you alluded to, has really only existed since the 1980's and there were no practical metrics to measure the phenomenon except using things like math to extrapolate them before that time. Unfortunately the naysayers in calling this unscientific are actually the ones not up to date on the science.
+TheHans .... regarding Ethan Winer, I have tremendous respect for him regarding room acoustics and standing wave control. With his published guidance, I was able to improve my studio tenfold! All I can say is that treating the primary reflection points and using a ton of bass traps will take your system to the next level (actually has a huge impact on SQ IMO) - though it can be unsightly! However, I believe he is incorrect when it comes to his interpretation of Nyquist. He is certainly more concerned with frequency balance (flat) vs timing balance of transients (minimized pre/post ringing).
"This is why I find double blind tests of audio not truly scientific to disprove high-resolution."
Double blind tests (assuming a proper execution) are the only scientific way to examine the real-world benefit of high resolution audio. I don't listen to music with an oscilloscope.
"To add to your point Hans, I remember dCS had a white paper on their website some time back, that explained what extreme band limiting in the digital domain did, and they showed it caused phase and amplitude shift, they had snapshots of it on oscilloscopes"
There is no phase-shift, if you use linear phase filters. That's basic knowledge of every electrical engineering student, you don't need mysterious special white papers for that.
"Humans can hear the harmonics of fundamental frequencies way above 20k(...)"
No, we can't hear beyond 20k. That's simply a wrong statement disproved by basically all science out there.
"(...)ringing artifacts smear in the time domain."
We can't hear the ringing, because its frequency is not within the hearable frequency range.
"If high order harmonics were not important then, for example, the type of piano or guitar you use and how it resonates would be inconsequential because they are all generating the same fundamental frequencies, but as any musician will tell you, it does matter."
There are usually enough harmonics below 20k to give instruments a unique sound. That not all guitars sound the same doesn't prove your claim.
"Simply put Nyquist is sufficient to prevent aliasing, but not sufficient to portray real life, and we are after it sounding like real life, not to just no longer have aliasing artifacts."
The illusion of real life starts where our sensory organs end. And that's at around 16-18k at least for my ears.
Nyquist only explains how to digitize a bandlimited signal. That our ears are bandlimited is not a result of Nyquist, but the foundation of its application.
Although you could say you lose information with quantization, this error only produces noise (quantization noise) which sets your maximum dynamic range, with no effect to the fidelity of the signal (provided dither is applied). This noise is 144dB below Full Scale at 24 bits, way below human hearing. Even at 16 bits and no dithering, the distortion produced by quantization is 100dB below Full Scale, and is in practice completely masked by a musical program at a sensible level.
Hoi Hans, het is al wetenschappelijk gemeten en bewezen wat de ruimtelijke-spatiale resolutie van de human hearing is, nml 1-2 graden. het heeft dus niets te maken met nyquist, maar de fase resolutie die nodig is om te weten waar een geluid vandaan komt, irrelevant voor mono, maar essentiel voor stereo, en de daarvoor nodige sampling rate, ver boven de 44.1k. Google het boek: Spatial Hearing, The psychoacoustics of human sound localization.
What are your thoughts regarding higher sampling rates (at least greater than 44.1) improving the imaging of the reproduced sound?
I have tried to express them in this video:-(
I think he is referring specifically to the effect on imaging that higher resolution audio may have vs. standard resolution (not necessarily MQA encoded). Improving the time domain with higher sample rates alone, may be advantageous to imaging, but how higher resolution will fix imaging (in particular) may not be obvious or part of the video above. Just saying. Again. Kudos on your awesome videos!
Thanks voor clearing this up. Trying to answer too many questions in too little time is claiming its toll, I am afraid. Indeed, the filtering artifacts do have a negative effect on imaging and placement. So reducing the artifacts does improve this. In what way and to what extend depends on where your started, what improvements were made , how these improvements were implemented and so on.
HI, I recently recorded guitar & vocal (Roland R-26, Digital Recorder omni + x/y mics) 24 Bit with 96 kHz Samplingrate instead 48 khz. Voc & git seemed to be the same, but noise and reverb are much better & more subtil in my ears - in the fast mix (Cubase 7 Elements).
Verry interesting vid, thanks for posting this, there is 1 more thing I wanna know. The more hertz the deeper the bass or the lower the hertz the lower the bass ? It is a bit confusing for me . I am looking for woofers to replase but I want woofers who can handle verry low basses. To what I schould look ??? high or low hertz ? thanks Hans
Low hertz/ 30 Hz is lower than 40 Hz.
@@TheHansBeekhuyzenChannel ik spreek engels tegen jou maar ik denk dat jij nederlands spreekt aan je naam te zien. Bedankt voor je berichtje terug . Dus hoe lager de HZ van je basswoofer is hoe lager je je woofer bassen kan produceren , heb ik dat juist begrepen ?
@@peterderidder2655 yep
@@TheHansBeekhuyzenChannel Thanks Hans
You might find this interesting because it supports high fidelity recording is superior to CD quality for those with good ears.
I have heard both sides of the debate and as a holder of a graduate degree in mathematics I am aware that many non mathematicians don't understand the language used in the theorems they quote and therefore infer incorrectly.
I could not actually obtain Shannon's actual work that consists of more than one theorem but I reconstructed what it must indicate and it is as follows:
A: As little as 2 samples in a complete sound wave will guarantee frequency.
B: Only 2 samples in a complete sound wave will not guarantee amplitude because there are an infinite number of curves that will fit the 2 samples.
C: To also guarantee amplitude you need at least 2 samples in a half cycle of the sound wave being recorded. Hence, ideally you would have 4 samples in the complete cycle or at a minimum 3 samples can suffice.
C above tells me that regular CD quality is going to be sketchy on reproduction for above roughly 14 KHz at best or even perhaps as low as 11 Khz. Why? Because while reproducing the frequency it is more or less guessing amplitude.
Another major issue is when a audible frequency has a ripple of high frequency variation superimposed across the entire cycle of the sound wave being recorded.
To a set of excellent ears, while perhaps not hearing the high frequency ripple due to its frequency being out side the range of hearing, they may process the sound like we process the brightness of a light that has the appearance of being dimmer due to pulse width modulation that has interupted the supply current to the light at a high frequency. However, what did the DAC do? It can't know what to do because it's data does not match a smooth sinewave and therefore the DAC must be following a set of instructions that produces something but not what was actually recorded. Does the actual sound wave sound identical to good ears? I doubt it thus I am reasonably certain that good ears under the right circumstances can detect analog from even digital sampled at 88 K.
You are right if we would consider short instantaneous sound byte of only one or two samples. But Shannon and Nyquist is about continuous signals and then they are correct. But, as said, it is impossible to bandwidth limit a signal as steep as done with CD without introducing time smearing.
@@TheHansBeekhuyzenChannel interesting for in my mind a "long duration of an actual acoustic sound" such as a single frequency of 8 Khz where each half cycle contains more than 2 samples
then
followed by a duration of 22Khz where each compete cycle found in the "interval of duration" had but 2 samples with & only with one sample in each half of each cycle in the sequence of data representing "the interval of duration" would pose a problem, because for any set of 2 samples indicating a complete cycle where the samples are separated by a known distance with one sample in the positive side of the wave and the other in the negative side of the wave the DAC has an infinite number of curves that will fit such a pair of (pos, neg) samples. What it then does seems to be an instruction coded in the DAC's software?
Granted what I am envisioning is overly simplified because we have algebraic addition of sound waves that then can be taken apart and seen as a Series of individual Sine and Cosine functions having unique coefficients in the series. Technically, it would be a transform as opposed to series but the series if sufficiently constructed accomplishes the same thing when infinite and darn close when the counter for the series such as N is as little as 14.
I would actually like to see Shannon's original work. Why? Because I believe it is at least one lemma and several Theorems and not a singular theorem as often indicated. The other reason is simply: the theorems must be constructed as an equivallance or strict implication meaning in the case of equivalence if A then B must be true and if B is True then A must be true. For the equivalence we simply have if A then B is True. Why it's important to see the actual theorem is that A in both of my examples if viewed as a hypothesis or proposition provides the boundary conditions that are required for B to be true. Mathematics is like ancient Greek in that if you do not know the language it is easy to infer the wrong meaning thus be mislead on what the theorem actually indicates. My speciality is probability and I can't begin to tell you how many people misread and then misuse the "the weak law of large numbers" as well as the "central limit theorem." Simply stated, if you don't speak Greek please don't translate the book for me and I am sure everybody understands my attitude and would feel the same too.
Wikipedia is your friend here.
@@TheHansBeekhuyzenChannel excellent, I will examine the Wikipedia article as a starter. Unfortunately, the writer of the article assumed the reader know's what all of the utilized notation stands in for such as Beta. This is often the case when it's becomes standard notation for the select few who specialize in the sub-field. I'll either derive representation or find a more complete accounting. Also, while sketching out the basis for the conclusion visually and in notation indicated at the end of the article ithe article does not completely provide the hypothesis needed for the conclusion and more suggests than explicitly indicates all that is sufficient but it's a good start none the less.
I am not doubting Shannon at all, rather I have been doubting what has been inferred by others because if I had a data stream of amplitude values (1,-1), (1, 1), (1,1), (1,-1)...repeat ad infinitum; while easily being able to ascertain frequency I would be hard pressed to tell you what specific sound of said frequency from the set of infinite possibilities would match the actual sound recorded. It could be a sound that warbled by changing amplitude every half cycle or one of constant amplitude during the entire duration of the actual sound.
Actually what I am talking about is what is referred to as Aliasing and Shannon and I are in agreement because his proof regarding this concludes with:
 and {\displaystyle x_{A}(t)} might not sound the same. But their samples (taken at rate fs) are identical and would lead to identical reproduced sounds; thus xA(t) is an alias of x(t) at this sample rate.
This is exactly what I am saying in that like in my example I have one representation from a given sampling rate but from conceivably two different sounds.
This is most interesting and very complex because it's not a single theorem rather several on the topic of digital representation and Shannon did it well. It will take a while to assemble all he did with complete understanding but it's more than meets the eye.
Well you were right Wikipedia was my friend. What I was pointing out that a mere 2 samples can lead to is called Aliasing. This is where two distinct actual sounds have the same digital representation and would play back from the recording exactly the same in spite of the originals being different.
Apparently when using 44K they apply filters to the input to prevent this. I suppose the filter simply removes the problematic high frequencies thus it no longer is faithful to the original. For me, I would up the rate to 88k so as to be faithful to the original but even then there are opportunities for excellent ears to hear a difference between analog and digital in my view.
The good news is your interest in High Fidelity above CD quality is real and not just in your imagination except you do need a brain that can differentiate small but important differences.
As far as Shannon, I was happy to find he and I were on the same page although I did not know of the term Aliasing until today. On the other hand his work on the topic of digital representation is more than most people seem to grasp and there is a lot I'd like to look at icontained in theorems other than the one pertaining to Aliasing. Who knows, maybe for giggles I will code my own DAC software to stay busy in the hour of isolation from Covid 19.
Thanks for pointing me towards the pages of info I needed to see and you have a great channel and don't let tin eared people say otherwise.
Mike
Nyquist is right
You are right.
I am right.
192khz is good for some applications.
44.1khz is mostly good enough for me.
I couldn't personally hear differences between the two rates on my studio Quality headphones. But I vaguely remember in my college, very steep filters having a very awkward graph. I agree with your theory on that as well.
Thanks!
I seriously don't know why all the people that claim it does or doesn't make a difference by going higher in sample rate, do not measure the differences with good measuring equipment.
There's a good audio blog I sometimes read that has. The blogger has done FIR tests, comparing different sample rates. Forgetting about all the filters he tested with, let's just take the industry standard linear phase filter:
5kHz @44.1kHz = no pre or after ringing
10kHz @44.1kHz = small pre and after ringing
--
5kHz @88.2kHz = no pre or after ringing
10kHz @88.2kHz = no pre or after ringing
20kHz @88.2kHz = small pre and after ringing
---
As you can see, the higher the sample rate, the further up in the frequency domain any ringing is pushed. Judging from this, the sample rate standards of 176.4kHz and 192kHz would completely rid the DAC of ringing within our audible range.
Now whether that is audible depends on an individual's hearing, equipment and environment.
I would say in terms of audibility, one could potentially hear the effects of ringing @44.1kHz and @48kHz as if I remember correctly, the actual Fourier ringing is double the impulse frequency (I may be slightly wrong). Judging from the data, this basically means that ringing could start around 8kHz and therefore add audible 16kHz+ noise to the audio. This would have the effect of making things like snares sound too sharp and artificially changing the soundstage.
When at 88.2kHz or 96kHz, there is technically still ringing but as the ringing probably kicks in around 16kHz then that would mean the added noise would be 32kHz+. This is not audible, but could still affect the sound. If the audio has a lot of higher treble energy, the speakers tweeter will be working overtime to reproduce not only the actual fundamental frequency but also its Fourier ringing too. I do not know how this would could affect the sound from an audible point of view.
In terms of audibility though, it also depends on the DAC's filter that is used. This is why the trend of having selectable filters exists, as DAC manufactures know that it makes a difference. Many DAC's can sound equal in tonality and dynamic range, but seemingly different in soundstage and attack. This is all to do with ringing.
Pre-ringing affects the transient response of upper frequencies, having a noisy response will add harshness or a dirtier attack. Having zero pre-ringing will keep the "attack" untouched and accurate, sounding natural and clean of a "digital sound".
Post-ringing affects the upper frequencies sense of space. Having a long post-ring will artificially increase the distance in between sounds, therefore making the soundstage bigger than it actually is. Having zero post-ringing will not stretch the soundstage and preserve its original distance and depth.
There is another thing you are not talking about that's the microphone, I don't know a mic with that higher frequency range spec. I google it and all I found mic's from 10-140 khz response.
Microphones that have such a high frequency range must be condenser mikes and have very small diaphragms. As a result they will have extremely poor signal to noise. Apart from that there is little to no content above 20 kHz in acoustical music and if there is, the energy is very low. That energy will get lost while traveling through the air. So in a concert hall it is unlikely frequencies above 20 kHz will reach you and if they do it is unlikely your hearing will register them.
I agree.
Goedenavond,
I must agree ... bit I am of the opinion that you cannot accurately describe a 20khz wave form with a 44.1Khz sample rate (you can perfectly describe a pure sine wave at 22,050 hz with a 44.1Khz sample rate ... so long as you know its a pure sine wave, but unless your sample rate is always exactly twice the frequency, you are going to loose information .... maybe variable sample rates are the answer, not sure how you'd poll that one off) So I also believe sample rates need to go up an order of magnitude to genuinely reproduce the sound.
Sample resolution (16 bit) I'm OK with, I don't feel there is much need to increase the bit depth of the signal after it has been mixed and processed at the studio and made available for public consumption, but I do think the sample rate needs to increase.
Anyways, the question I wanted to ask you .....
it is kind of old technology, and forgive me if you have explained it in another video (I'm sure if this is the case another follower of your channel will respond to this post with a link) ... 1 bit dac ....way back in the dark old days (when cd players were mostly 14 bit pretending to be 16 bit) there emerged '1-bit dac' players ... I have a Mission/Cyrus CD Player from the dark old days that boasts '1 bit dac' .... now I've been programming computers, and have been doing so since early 80's, and I know you cannot simply break 16 bits into 16 single bits and process each bit without consideration of the power of that bit. and if 1-bit dacs are truly that, then they are only capable of either outputting +v or 0v, and nothing in between.
So .... can you explain (demystify) what is 1-bit DAC, and how does (how did) it work, because I've never figured it out.
dank je wel
Best wishes
Duncan
You can perfectly describe a 20 kHz waveform using a 44.1 kHz sampling frequency as anyone with a scope can show you. You can not describe perfectly a 20 kHz square wave at the same sampling frequency since that is built from all sine waves from DC to eternal frequency. Since the signal is filtered at 20 kHz, this is impossible. As it is impossible to record that square wave on an analogue tape recorder since that is band limited by design too. 1 Bit audio converters are comparable to what is used in DSD technology. Watch DSD Explained at czcams.com/video/hXFIq11JAas/video.html
A very interesting video (and part 2), thank you
thank you, Hans, for reinforcing my belief about the limitation of the Nyquist theorem. i have been a proponent of higher sampling rates and was an early supporter of SACD although i get flack from some about the supposed noise in the system. what i find is the relaxation/satisfaction factors that i feel are consonant with analog recording.
i have heard some superior playback of Peter Mc Grath's digital rcordings on Constellation electronics and Wilson Audio speakers at Audio Salon in Santa Monica CA.
i have watched a number of your CZcams presentations and have always found you more than credible.
...hifitommy
Interesting video, I am finding that you are right on most of the things you are saying. However there are a few things that you either have misinterpreted or, as you mentioned your self "not clever enough to understand" or as I would say, possibly not having the theoretical foundation to understand. First of all, the Nyquist-Shannon theorem is mathematically proven. The implementation of that theory is something else. So, the theorem does not "fail to successfully bandlimit a signal". The theorem has many applications beyond sound. If for instance I have a knob that sends information in "steps", then the receiver must have adequate time resolution to "read" these changes from the knob without creating aliasing artifacts. Another application is video. If you live in Europe and you shoot video using typical lamps, then your shutter rate, according to Nyquist Shannon must be multiples of 50Hz (the AC frequency) so you won't have flickering. So, the theorem helps to get around situations like these. On another note, the practical application of increasing the sampling frequency is not going to offer much value to the every day use of audio technology and rather cause problems. Those who can understand the differences in DACs are a very very small percentage of people that use sound for any reason. Increasing the sampling rate and using longer FIR filters to bandlimit signals means longer latencies, something that will cause problems in live and production sound application, will increase exponentially the amount of resources a computer might need to process realtime audio, at the same fashion, this inevitably will affect video as well and in the end whatever the process of creating audio, this audio will be consumed by people on cars, headphones, earpods or whatever less than perfect audio reproduction system. Moreover, the mp3 revolutionised the music industry as it made possible to share sound over the internet... yeah, piracy skyrocketed, but the system adapted to spotify, youtube etc, and will keep adapting. So, it doesn't make sense to have a higher resolution audio file, that the extra resolution is used to bandlimit better, and then make it mp3... there's not an infrastructure yet that would allow us to stream lossless sound over the internet in the scale that we do today, so.. an idea like that, it will of course be rejected. So, why put the effort on something that will not benefit the majority of the people that use it? If there's a specific reason for you or anyone that uses sound for research, your logic is absolutely sound (pun intended) and they should seriously consider it. However, such change would mostly affect the public negatively up to the point that the technology catches on. On the other hand, most people barely can listen up to 18kHz due to noise exposure for recreation, work or simply a loud place to live. I don't see anyone getting worried about increasing resolution anytime soon.
I merely ment to say that Nyquist and Shannon might have developed a theorem that is still impossible to implement
It seems like a better way to record audio is with a high-quality analogue medium rather than digital after watching some of your videos. If there was a method that didn't degrade that is, as copys would still degrade it.
analogue recorders have their losses too. For instance, due to the so called 'head bump' you never get as tight a bass as with digital. And there is more.
Hans, ter info: ik wilde je geschreven artikel delen, maar je website lijkt corrupt te zijn.
Klopt. Ik zoek naar een oplossing
Thanks so much for keeping matters logical and scientific. I'd like to understand why the artifacts mentioned from sampling exist. Is there a simple example like there is with sampling images eg aliasing or maybe others. Perhaps you have a reference that talks about filters in both time and frequency domain with illustration as can be done with image DSP. It's kind of easy to understand 4k TV vs 8k TV as it's ultimately visual but very hard to relate to 96k and above audio sampling in a physical way... Also I wonder as an example the philosophy behind the design of filters in dragonfly cobalt and high end chord dac how do they find the filters that make it so good and how much work is needed. Is the work art or science and why do I have to pay 3K up for a DAC !
I have already covered a number of items you mention. Just shop around in my playlist page: czcams.com/channels/R4tuhqPppVp-PD0q17sPEA.htmlplaylists?view=1&sort=dd&shelf_id=0
Hello, i think that your explanation of the NOS tecnology was a little to short, any chance for you to go deeper?
Do you have any listening experience with later converters?
thank's
Francesco
why do we need to attenuate high frequency content on playback if it is already attenuated in recording/mixing?
good question. I might just make a video on this. Can't do it in a short answer., sorry.
I read up on wikipedia. It is because a frequency component in the discrete domain can be represented with a combination of multiple frequencies in the continuous domain. Low pass filtering on reconstruction increases the likelihood that the frequency components is represented correctly. Eg. 20 Hz in the discrete domain with a sampling rate of 48KHz can be represented as 20, 47980, 48020Hz, etc in the continuous domain. Oversampling, filtering and mechanical limitations are what makes 20Hz discrete really become 20Hz continuous.
Yes, but do you understand it now?
Somewhat, I don't know how the spurious frequencies spawn in the first place but I just assume it is a phenomenom in physics.
You might want to watch czcams.com/video/z8dIzTZaRFY/video.html
Pulses are not real world. Sine waves are.
I agree completely. But digital audio doesn't reproduce pulses but pure wave forms. Just connect a scope to the output of a modern DAC or CD-player and watch a sine wave. You will see it does output a sine wave and not pulses. The misunderstanding comes from a simplified explanation of how digital works.
Perhaps viewing this video will help: czcams.com/video/z8dIzTZaRFY/video.html
:)
Most of sound you hear in music is not sine wave! Sure you can consider translating a time domain cymbal pressure wave mathematically in the frequency domain and say it is the superposition of this sine waves at this frequencies with this respective intensities, nobody is stopping you, but oh boy, look at the harmonic complexity of a note played on a violin, although the player might try to keep it periodical it's way far from a sine wave! Forget about a snare, that's closer to an impulse.
Klipsch Engineering That’s why Fourier transformations are employed for analysis. Sine waves are like Epicycles, given enough of them anything can be constructed..
Doe je goed, Hans! Mooie uitleg!
Thank you.
Thank you Sir! Most, just love to learn new information... Keep it coming.
"It's only a theory" .. so is gravity 🙈
It's actually a theorem, not a theory. Gravity is something that exists, sampling is invented. With the remark I meant that the theorem can't be implemented perfectly in real life. Limiting bandwidth without artefacts in time is impossible and our hearing is far more sensitive to timing than to frequency.
In love listening to this guy, and trying to keep up!! Smarter than me, he is..
I want to understand how time resolution works in a analogue system say a vinyl. Does a 20khz single from a vinyl and flac over pcm have the same time resolution? And if analogue is continuous without any interval then shouldn’t it be infinitely superior to digital? Also maybe u can make a video elaborating on time resolution. How does it affect the sound etc. thank you
Perhaps this helps: czcams.com/video/nFWYM_ZQN2Y/video.html
Great video... yes, I understand and agree with your points and I too, enjoy the music ;-)
Your written english version only shows the first chapter.
My question is in regard to playback versus initial sampling. Let me give an example. I record a simple session at 96/24 of an acoustic guitar and a vocal using two mic's. Once it is finished, I do a finalized stereo bounce down into an mp3 file so I can play it on a digital medium that can only play mp3 files. I also do the exact same thing again, except this time when recording the original tracks, I use 48/24 instead of 96/24, and it just doesn't sound as good as the session that was done at 96/24 then bounced down to a stereo mp3 file. Both sessions ultimately end up in a mp3 stereo file, so why does the one tracked at 96/24 sound better than the one tracked at 48/24?
MP3 is a lossy way of storing music. The higher the quality of the source material, the easier it is for the MP3 encoder to decide what part of the information hurts the least when left out. It's the same with pictures. If you want a high quality, low file size internet photo, start with a very high quality losslessly encoded photo at very high resolution.
Thanks for your response. I am currently looking into Antelope Audio products because they seem to be the one of the only pro audio component producers that focus on making their products with 192 khz capability at an affordable price.
A very good point - I watched this initially thinking "this will be a bunch of nonsense" but actually, everything you're saying I can't argue with - at least on an oscilloscope. I'm very skeptical that low pass filter artifacts are going to have an audible impact on the signal in a double-blind test, but I appreciate your video and the fact you based it on sound science. Anyone up for arranging a double-blind test? :)
Hearing the artefact differences between a bad and a very good product is rather easy. The difficulty for you is to set up such a comparison since it might be difficult to get the right equipment. In DAC's predominantly two kinds of artefacts cause clear sound degradation: jitter and the reconstruction filter/upscaling filter. So you need proper low jitter in both DAC's to compare only the filtering. About double blind testing audio, watch this: czcams.com/video/QG6LS9VDZlg/video.html
👍👍👍👍👍Nothing to add....
🙏🏻
Amazing video
Is this somewhat connected to bad or low quality DACs question that I have sometimes wondered?
If one routes 44kHz CD-signal digitally to low quality DAC for example in cheap av amp. It will sound in my opinion worse. If one uses same cheap av amp as only amplifier and feeds analog signal from high quality CD player (which has better DACs) It will sound much better.
I've allways wondered why this is? Even though digital signal is the same and bit is a bit. Is this issue connected to DAC or DAC output filters?
I really can't say
Yes there can be many other factors in play. I just wondered when I saw this video. That if it would be the DAC output filters that cause the difference.
Then it would simple way to improve audio by running the digital signal at higher sampling frequency. With higher sampling the output filter is not that critical because it would be out of hearing range, right? And so we would get more out of low end systems, maybe?
This would be interesting to test. I unfortunately don't have equipment to do so.
An excellent explanation, thank you
🙏🏽
Thank you so much for this useful video. I have heard many different dacs from different manufacturers using different chips to do the job and i always preferred the NOS implementations. They somehow sound more natural to me in comparison, so i am using one in my own system.
Well, see this video: czcams.com/video/EykkzxUstL4/video.html
So 2L's DXD should be even better and offer less time smearing?
In theory at least.
@@TheHansBeekhuyzenChannel Thanks, Hans. Like Stephen Weiss I'm really sorry that you have been verbally abused for your opinions concerning higher sampling rates, which I wholeheartedly share. High-end digital audio isn't cut and dried like some people tend to believe. It's very complicated with so many factors contributing to its success.