Home |
Search |
Today's Posts |
#1
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
Looking through a recent posting of papers given at the most recent AES
meeting in New York (Thanks Scott!), I found a mention of the following paper that sheds some light on other listening tests that have shown that, all other things being equal, 24/96 sampling isn't necessarily better-sounding than 24/44 sampling. Audio Engineering Society Convention Paper 5876: Perceptual Discrimination between Musical Sounds with and without Very High Frequency Components This paper describes the test methodology and the results of a series of listening tests performed by researchers at NHK Science & Technical Research Laboratories, Tokyo, Japan. These tests compared the playback of recordings with and without audio signals above 21 KHz. 19 different musical selections and one synthetic sound were used: 1 "Satsuma-Biwa" "Satsuma-Biwa" 2 Litha Drums, Bass, Pf (Jazz Piano Trio) 3 Meditation Vn, Pf 4 Romanian Folk Dances Vn, Pf 5 Intermezzo de "Carmen" Fl, Pf 6 Beethoven: Sym. No.9 4th Mov. Picc 7 Bach: Suite for Vc No.2 - Prelude Sax 8 Bach: Suite for Vc No.6 - Prelude Sax 9 Piece en forme de Habanera Sax, Pf 10 Partie Sax, Pf, Perc 11 Sednalo Bulgarian Chorus (SACD ARHS-1002) 12 TihViatar Bulgarian Chorus (SACD ARHS-1002) 13 Meditation+White Noise Vn, Pf, High frequency band consists of only white noise. 14 Airs Valagues Fl, Pf 15 Tchaikovski: Sym. No.6 3rd Mov. Full Orchestra 16 Doralice Vo, Gt (Bossa Nova) 17 Chega de Sauadade Vo, Gt, Pf, Perc (Bossa Nova) 18 tiny rose Vo, Pf, Gt, Fl, Perc ("the birds") 19 butterfly Vo, Pf, Gt, Perc ("the birds") 20 Autumn Leaves Drums, Bass, Pf (Jazz Piano Trio) Notably 2 SACD selections were used. "First, 36 subjects evaluated 20 kinds of stimulus, and each stimulus was evaluated 40 times in total. The results showed no significant difference among the sound stimuli, but that the correct response rate for three sound stimuli was close to the significance probability (5% level). It is concluded that one subject attained a 75% correct response rate which constituted a significant difference. In order to make a strict statistical test, we conducted a supplementary test with this subject who had attained the best correct answer rate in the first test. This subject evaluated six kinds of sound stimulus, and then evaluated each sound stimulus 20 times. As a result, no significant difference was found among the sound stimuli, and so this subject could not discriminate between these sound stimuli with and without very high frequency components." In other words, of 36 listeners, only one listener scored substantially better than random guessing, and when retested, he could not duplicate his earlier results. This indicates that his results were due to luck. A study of statistics and actual experience suggests that with a group of 36 listeners, it is pretty much certain that one or more listeners will get good scores due to luck, and that they won't be able to duplicate those results when re-tested. So, you can flip pennies or compare 24/44 to 24/96 and get pretty much the same results, provided you hold all other relevant variables equal. |
#2
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
They should maybe try a similar listening experiment comparing 24 tracks
of program at the various sampling rates, summed through an analog mixing bus. Maybe I can't tell the difference between 44.1 and 48k sampling with two tracks, but I have been able to tell between when working with a Sony 3348 on a daily basis when you have a lot of tracks up. Will Miho NY Music & TV Audio Guy Fox News/Fox & Friends "The large print giveth and the small print taketh away..." Tom Waits "Arny Krueger" Looking through a recent posting of papers given at the most recent AES meeting in New York (Thanks Scott!), I found a mention of the following paper that sheds some light on other listening tests that have shown that, all other things being equal, 24/96 sampling isn't necessarily better-sounding than 24/44 sampling. Audio Engineering Society Convention Paper 5876: Perceptual Discrimination between Musical Sounds with and without Very High Frequency Components This paper describes the test methodology and the results of a series of listening tests performed by researchers at NHK Science & Technical Research Laboratories, Tokyo, Japan. These tests compared the playback of recordings with and without audio signals above 21 KHz. 19 different musical selections and one synthetic sound were used: 1 "Satsuma-Biwa" "Satsuma-Biwa" 2 Litha Drums, Bass, Pf (Jazz Piano Trio) 3 Meditation Vn, Pf 4 Romanian Folk Dances Vn, Pf 5 Intermezzo de "Carmen" Fl, Pf 6 Beethoven: Sym. No.9 4th Mov. Picc 7 Bach: Suite for Vc No.2 - Prelude Sax 8 Bach: Suite for Vc No.6 - Prelude Sax 9 Piece en forme de Habanera Sax, Pf 10 Partie Sax, Pf, Perc 11 Sednalo Bulgarian Chorus (SACD ARHS-1002) 12 TihViatar Bulgarian Chorus (SACD ARHS-1002) 13 Meditation+White Noise Vn, Pf, High frequency band consists of only white noise. 14 Airs Valagues Fl, Pf 15 Tchaikovski: Sym. No.6 3rd Mov. Full Orchestra 16 Doralice Vo, Gt (Bossa Nova) 17 Chega de Sauadade Vo, Gt, Pf, Perc (Bossa Nova) 18 tiny rose Vo, Pf, Gt, Fl, Perc ("the birds") 19 butterfly Vo, Pf, Gt, Perc ("the birds") 20 Autumn Leaves Drums, Bass, Pf (Jazz Piano Trio) Notably 2 SACD selections were used. "First, 36 subjects evaluated 20 kinds of stimulus, and each stimulus was evaluated 40 times in total. The results showed no significant difference among the sound stimuli, but that the correct response rate for three sound stimuli was close to the significance probability (5% level). It is concluded that one subject attained a 75% correct response rate which constituted a significant difference. In order to make a strict statistical test, we conducted a supplementary test with this subject who had attained the best correct answer rate in the first test. This subject evaluated six kinds of sound stimulus, and then evaluated each sound stimulus 20 times. As a result, no significant difference was found among the sound stimuli, and so this subject could not discriminate between these sound stimuli with and without very high frequency components." In other words, of 36 listeners, only one listener scored substantially better than random guessing, and when retested, he could not duplicate his earlier results. This indicates that his results were due to luck. A study of statistics and actual experience suggests that with a group of 36 listeners, it is pretty much certain that one or more listeners will get good scores due to luck, and that they won't be able to duplicate those results when re-tested. So, you can flip pennies or compare 24/44 to 24/96 and get pretty much the same results, provided you hold all other relevant variables equal. Will Miho NY Music & TV Audio Guy Off the Morning Show! & sleepin' In... / Fox News "The large print giveth and the small print taketh away..." Tom Waits |
#3
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44sampling
Arny Krueger wrote: This paper describes the test methodology and the results of a series of listening tests performed by researchers at NHK Science & Technical Research Laboratories, Tokyo, Japan. These tests compared the playback of recordings with and without audio signals above 21 KHz. Arny, could you tell us what the reproduction chain was? Bob -- "Things should be described as simply as possible, but no simpler." A. Einstein |
#4
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
"Arny Krueger" wrote
These tests compared the playback of recordings with and without audio signals above 21 KHz. Who cares about what is happening above 20K? The critical difference is what happens at frequencies you can hear. At 11 Khz you have only 4 samples per cycle using 44.1Khz sampling frequency but nearly 9 samples per cycle using 96Khz. Anthony Gosnell |
#6
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 samplin
"Mike Rivers" wrote
writes: At 11 Khz you have only 4 samples per cycle using 44.1Khz sampling frequency but nearly 9 samples per cycle using 96Khz. And this means? If you're suggesting that the 11 kHz sine wave will be more accurately reproduced from 9 samples than from four, you're wrong, provided that all other rules of sampling have been followed. Since when did music consist only of pure sine waves? -- Anthony Gosnell to reply remove nospam. |
#7
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
"Bob Cain" wrote in message
Arny Krueger wrote: This paper describes the test methodology and the results of a series of listening tests performed by researchers at NHK Science & Technical Research Laboratories, Tokyo, Japan. These tests compared the playback of recordings with and without audio signals above 21 KHz. Arny, could you tell us what the reproduction chain was? Unfortunately, the test system is only described with a diagram, and of course this is a text-only forum. However, I'll try to crib a few captions: DAW SADiE ATEMIS Cool Edit Pro D/A Dcs 954 Master Clock dcs 992 Controller Laguna Hills SYSTEM 1000E Amp. SONY FA777ES Super Tweeter PIONEER PT-R9 Power Supply Accuphase PS-1200V Speaker B&W Nautilus 801 Amp. Marantz PA02 I get the impression that there were two separate, independent reproduction chains, one for 21 KHz and one for 21 KHz. This was no doubt done to minimize intermodulation. I suspect they did the 21 KHz filtering with Cool Edit Pro and used Cool Edit's multitrack facilites to handle the playback. I'm a little confused because I'm under the impression that the Sadie Atemis workstation is Mac-based however it does exchange data with PCs. The 21 KHz reproduction chain used a DCS 954 DAC, a Sony FA 777ES amp, and a Pioneer PT-R9 super tweeter. The 21 KHz reproduction chain used a DCS 954 DAC, a Marantz PA02 amp, and a B&W Nautilus 801 speaker system. The DCS 992 handled clocking for both the low and high frequency DACs. It was also stated that the listening room conformed to IEC recommendation BS 1116-1 which is very stringent. For example, BS 1116-1 sates that under no circumstances should the background noise exceed NR 15. |
#8
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
On Sat, 15 Nov 2003 14:40:03 +0200, "anthony.gosnell"
wrote: "Arny Krueger" wrote These tests compared the playback of recordings with and without audio signals above 21 KHz. Who cares about what is happening above 20K? The critical difference is what happens at frequencies you can hear. At 11 Khz you have only 4 samples per cycle using 44.1Khz sampling frequency but nearly 9 samples per cycle using 96Khz. Anthony Gosnell I can't tell if you are agreeing or disagreeing with Arny. If you're saying that 44.1 has an inaccurate representation of higher frequency information, sure, but that inaccuracy shows up mainly as overtones above 20K that's removed by the Nyquist filter. The rest is what's called quantization noise, and dithering takes care of that at the expense of a tiny amount of white noise too small to bother about. |
#9
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
"Arny Krueger" wrote in message ...
"It is concluded that one subject attained a 75% correct response rate which constituted a significant difference. In order to make a strict statistical test, we conducted a supplementary test with this subject who had attained the best correct answer rate in the first test. Why is it necessary to conduct a supplemental test only with the subject who scored the highest? Shouldn't the researchers have retested all subjects equally? If he had one "lucky" run wouldn't it have been equally possible that another of the subjects had an "unlucky" run and would have scored higher in subsequent tests? |
#10
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44sampling
Arny Krueger wrote: The 21 KHz reproduction chain used a DCS 954 DAC, a Sony FA 777ES amp, and a Pioneer PT-R9 super tweeter. The 21 KHz reproduction chain used a DCS 954 DAC, a Marantz PA02 amp, and a B&W Nautilus 801 speaker system. The DCS 992 handled clocking for both the low and high frequency DACs. Hrmmph. If they used different repro chains I don't see how any valid conclusions at all about intrinic differences due to sample rate can be drawn from the listening tests. Thanks for the info. Bob -- "Things should be described as simply as possible, but no simpler." A. Einstein |
#11
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
"anthony.gosnell" wrote in message
"Arny Krueger" wrote These tests compared the playback of recordings with and without audio signals above 21 KHz. Who cares about what is happening above 20K? Good question. The critical difference is what happens at frequencies you can hear. Only if there is a meaningful difference. At 11 Khz you have only 4 samples per cycle using 44.1Khz sampling frequency but nearly 9 samples per cycle using 96Khz. What difference would you expect this to make? Digital theory and practice say that it takes only slightly over 2 samples per cycle to get as good of a sampling job of a sine wave as you can imagine. So 4 samples per cycle is overkill and 9 samples per cycle is gross overkill. IOW, there's no meaningful difference in accuracy. |
#12
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
On Sat, 15 Nov 2003 19:24:04 -0500, "Arny Krueger"
wrote: Digital theory and practice say that it takes only slightly over 2 samples per cycle to get as good of a sampling job of a sine wave as you can imagine. So 4 samples per cycle is overkill and 9 samples per cycle is gross overkill. IOW, there's no meaningful difference in accuracy. Isn't this only true for perfect, non-quantized samples? Chris Hornbeck new email address "That is my theory, and what it is too." Anne Elk |
#13
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 samplin
"anthony.gosnell" wrote in message
"Mike Rivers" wrote writes: "anthony.gosnell" wrote in message At 11 Khz you have only 4 samples per cycle using 44.1Khz sampling frequency but nearly 9 samples per cycle using 96Khz. And this means? If you're suggesting that the 11 kHz sine wave will be more accurately reproduced from 9 samples than from four, you're wrong, provided that all other rules of sampling have been followed. Since when did music consist only of pure sine waves? Anthony, since you are the author of the example based on sine waves, that would be your question to answer. |
#14
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
"Chris Hornbeck" wrote in message
On Sat, 15 Nov 2003 19:24:04 -0500, "Arny Krueger" wrote: Digital theory and practice say that it takes only slightly over 2 samples per cycle to get as good of a sampling job of a sine wave as you can imagine. So 4 samples per cycle is overkill and 9 samples per cycle is gross overkill. IOW, there's no meaningful difference in accuracy. Isn't this only true for perfect, non-quantized samples? Please explain what you mean by that. |
#15
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
On Sat, 15 Nov 2003 23:08:05 -0500, "Arny Krueger"
wrote: "Chris Hornbeck" wrote in message On Sat, 15 Nov 2003 19:24:04 -0500, "Arny Krueger" wrote: Digital theory and practice say that it takes only slightly over 2 samples per cycle to get as good of a sampling job of a sine wave as you can imagine. So 4 samples per cycle is overkill and 9 samples per cycle is gross overkill. IOW, there's no meaningful difference in accuracy. Isn't this only true for perfect, non-quantized samples? Please explain what you mean by that. Sorry, I guess that is about as clear as mud. How about: Isn't this only true for samples of infinite wordsize? Chris Hornbeck new email address "That is my theory, and what it is too." Anne Elk |
#16
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
"Chris Hornbeck" wrote in message
On Sat, 15 Nov 2003 23:08:05 -0500, "Arny Krueger" wrote: "Chris Hornbeck" wrote in message On Sat, 15 Nov 2003 19:24:04 -0500, "Arny Krueger" wrote: Digital theory and practice say that it takes only slightly over 2 samples per cycle to get as good of a sampling job of a sine wave as you can imagine. So 4 samples per cycle is overkill and 9 samples per cycle is gross overkill. IOW, there's no meaningful difference in accuracy. Isn't this only true for perfect, non-quantized samples? Please explain what you mean by that. Sorry, I guess that is about as clear as mud. How about: Isn't this only true for samples of infinite wordsize? Only if you want absolutely perfect results! If you have finite word size then you have finite SNR. The finite SNR creates ambiguities in how precisely the sine wave has been measured. |
#17
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 samplin
A large part of the problem is that to materialize these perfect sine waves
(up to the Nyquist frequency of 1/2 the sample rate) you need some pretty severe reconstruction filters. These real-world filters have audibly destructive effects on audio in the passband and, more significantly, on impulse response. A perfectly reconstructed sine wave and good sound are far from the same thing. Using double and quadruple sample rates moves filter artifacts further away from the audible frequency range. That is the advantage of over-sampling, rather than a frequency response up to 50 or 100kHz. |
#18
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 samplin
"Arny Krueger" wrote in message
... "anthony.gosnell" wrote in message "Mike Rivers" wrote writes: "anthony.gosnell" wrote in message At 11 Khz you have only 4 samples per cycle using 44.1Khz sampling frequency but nearly 9 samples per cycle using 96Khz. And this means? If you're suggesting that the 11 kHz sine wave will be more accurately reproduced from 9 samples than from four, you're wrong, provided that all other rules of sampling have been followed. Since when did music consist only of pure sine waves? Anthony, since you are the author of the example based on sine waves, that would be your question to answer. Arny, I didn't say anything about sine waves. I just said "At 11 Khz", you were the one who assumed that this must of course be a sine wave, and so I challenged your assumption. With just four samples per cycle you can reproduce that frequency but you don't really stand a chance at getting the shape right. -- Anthony Gosnell to reply remove nospam. |
#20
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 samplin
anthony.gosnell wrote:
Since when did music consist only of pure sine waves? All waveforms can be decomposed into pure sound waves. --scott -- "C'est un Nagra. C'est suisse, et tres, tres precis." |
#21
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 samplin
"Remixer" wrote in message
A large part of the problem is that to materialize these perfect sine waves (up to the Nyquist frequency of 1/2 the sample rate) you need some pretty severe reconstruction filters. A major advance came about a decade ago when these filters were finally moved into the digital domain with great success. It took about ten years of fits and starts to get things REALLY right. These real-world filters have audibly destructive effects on audio in the passband They may or they may not have audibly destructive effects. For example, I've tested fine brick-wall filters that have less phase shift up to 20 KHz and beyond, than a fine power amp. I've posted links to the phase/amplitude test results here many times. and, more significantly, on impulse response. The bottom line is your claim that "These real-world filters have audibly destructive effects on audio in the passband". IME you missed an important hedge-word, namely the word *can*. The correct statement is: "These real-world filters can have audibly destructive effects on audio in the passband". That means they may or they may not have audibly destructive effects. In the end it all comes down to listening. The only valid way to listen for potentially audible artifacts in good converters is the canonical level-matched, time-synchronized, bias-controlled listening test. However there are many perfectly acceptable ways to do good listening tests as long as these three requirements are paid attention to. For three years I've posted *everything* it takes to do a good listening tests of a number of converters at www.pcabx.com, except a DAW and a good monitoring system. If you're in this game, then you have those two remaining ingredients. If you try the PCABX web site listening tests you will find that some converters audibly trash sound quality in just one pass, and others don't make *any* audible changes after 10 or 20 passes. AFAIK, nobody who has ever tried the same thing by other reasonable means has produced results that significantly differ. A perfectly reconstructed sine wave and good sound are far from the same thing. Agreed. But we've had at least a few very good converters for years, They don't make any audible changes to even the most complex, demanding musical sounds. Let your ears (and just your ears) be your guide! Using double and quadruple sample rates moves filter artifacts further away from the audible frequency range. If a good 44.1 KHz converter has no audible artifacts with demanding and complex sounds, even after the conversion process is repeated 10 or 20 or even 40 times, what audible artifacts are we talking about anyway? That is the advantage of over-sampling, rather than a frequency response up to 50 or 100kHz. If it's not broke, don't fix it! |
#22
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 samplin
"anthony.gosnell" wrote in message
"Arny Krueger" wrote in message ... "anthony.gosnell" wrote in message "Mike Rivers" wrote writes: "anthony.gosnell" wrote in message At 11 Khz you have only 4 samples per cycle using 44.1Khz sampling frequency but nearly 9 samples per cycle using 96Khz. And this means? If you're suggesting that the 11 kHz sine wave will be more accurately reproduced from 9 samples than from four, you're wrong, provided that all other rules of sampling have been followed. Since when did music consist only of pure sine waves? Anthony, since you are the author of the example based on sine waves, that would be your question to answer. Arny, I didn't say anything about sine waves. I just said "At 11 Khz", you were the one who assumed that this must of course be a sine wave, and so I challenged your assumption. The 11 KHz number is yours. In a 44 KHz sampled system at 11 KHz and above there are nothing but sine waves and combinations thereof. When you say "cycle" Anthony, you are limiting your discussion to periodic waves. Any periodic wave can always be thought of as being a linear combination of sine waves. Any wave that has been brick-wall filtered at 22 KHz contains no sine wave components that are above 22 KHz. A sine wave at any frequency can be fully characterized by its frequency, amplitude and phase. It takes slightly more than two data points to fully determine the frequency, amplitude and phase of a sine wave. These are mathematical theorums and corolaries that are over 170 years old and have stood the test of time. When I say "slightly more than 2" I'm referring to 2 plus a mathematical delta, the smallest amount that can be conceived of. Therefore, slightly more than two sine waves per 22 KHz cycle is sufficient to fully characterize any wave that has been brick-wall filtered at 22 Khz. With just four samples per cycle you can reproduce that frequency but you don't really stand a chance at getting the shape right. As soon as *any* signal is brick-wall filtered at 22 KHz, there is no shape that can't be gotten right with a tiny bit more than four samples per cycle at 11 KHz or above. For more information please read the rec.audio.pro faq, particularly Question 5.12 "How can a 44.1 kHz sampling rate be enough to record all the harmonics of music?" |
#23
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
This paper describes the test methodology and the results of a series of
listening tests performed by researchers at NHK Science & Technical Research Laboratories, Tokyo, Japan. These tests compared the playback of recordings with and without audio signals above 21 KHz. Ok, has everyone lost the plot here or is it just me? So, what this test says is that basically music sounds the same regardless of whether or not it has everything above 21khz removed. So what the hell has this got to do with 24/48 or 24/96 recording? Nothing. Here is the simple computer programmers explanation of sampling rates. Analog sound is converted to digital. In order to do this it needs to be stored as bits. The more bits we can use to re-create the analog wavesform, the better the sound. So, the analog waveform is converted to digital in the AD process and expressed as a series of numbers. The waveform is sampled at the sample rate, like 48000 times a second or 96000 times a second. Each of these samples is then stored as a number. The the precision of this number is determined by bit size. A 16 bit number is significantly smaller and therefore less precise than a 24 bit number. So, in a nutshell. Moving from 16 bit to 24 bit, we have 8 extra bits per sample to represent the analog wave which is a massive gain. Moving from 48khz to 96khz we simply double the number of bits we have. Which is why people will easily notice 16 bit vs 24 bit and less so 48khz vs 96khz. Nothing to do with the peak human hearing frequency of 20khz as expressed in the start of this thread. |
#24
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
On 16 Nov 2003 06:14:51 -0800, (mike rogers)
wrote: This paper describes the test methodology and the results of a series of listening tests performed by researchers at NHK Science & Technical Research Laboratories, Tokyo, Japan. These tests compared the playback of recordings with and without audio signals above 21 KHz. Ok, has everyone lost the plot here or is it just me? So, what this test says is that basically music sounds the same regardless of whether or not it has everything above 21khz removed. So what the hell has this got to do with 24/48 or 24/96 recording? Nothing. Here is the simple computer programmers explanation of sampling rates. Analog sound is converted to digital. In order to do this it needs to be stored as bits. The more bits we can use to re-create the analog wavesform, the better the sound. So, the analog waveform is converted to digital in the AD process and expressed as a series of numbers. The waveform is sampled at the sample rate, like 48000 times a second or 96000 times a second. Each of these samples is then stored as a number. The the precision of this number is determined by bit size. A 16 bit number is significantly smaller and therefore less precise than a 24 bit number. So, in a nutshell. Moving from 16 bit to 24 bit, we have 8 extra bits per sample to represent the analog wave which is a massive gain. Moving from 48khz to 96khz we simply double the number of bits we have. Which is why people will easily notice 16 bit vs 24 bit and less so 48khz vs 96khz. Nothing to do with the peak human hearing frequency of 20khz as expressed in the start of this thread. Clearly being a simple programmer isn't sufficient. Moving from 16 to 24 bits improves matters if - and only if - the noise floor of the original analogue signal is below that of the 16 bit dither signal. The chances of that happening in any real recording are vanishingly close to zero. In virtually any scenario encountered in real life, 16 bits record just as high a quality as 24. There may well be special demo recordings that don't obey this rule of thumb. As regards sampling rate, the situation is perhaps not quite as clear. Certainly it is possible to find microphones that are pretty flat up to 20kHz, and have useful output above. Whether that makes any audible difference to a recording is debatable. Certainly recording above 20kHz will result in the capture of some pretty major untreated resonances from any microphone - manufacturers aren't quite as fussy about flatness up there. If you accept that 20kHz represents a useful upper limit to human hearing, and there is nothing significant above, then 44.1 is every bit as good as 48 or 96. 44.1 captures *everything* up to 20kHz with no exceptions. d _____________________________ http://www.pearce.uk.com |
#26
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
mike rogers wrote:
This paper describes the test methodology and the results of a series of listening tests performed by researchers at NHK Science & Technical Research Laboratories, Tokyo, Japan. These tests compared the playback of recordings with and without audio signals above 21 KHz. Ok, has everyone lost the plot here or is it just me? So, what this test says is that basically music sounds the same regardless of whether or not it has everything above 21khz removed. So what the hell has this got to do with 24/48 or 24/96 recording? Nothing. It does, in that the only thing that the higher sample rate buys you is the ultrasonic response. Here is the simple computer programmers explanation of sampling rates. Analog sound is converted to digital. In order to do this it needs to be stored as bits. The more bits we can use to re-create the analog wavesform, the better the sound. Not necessarily, no. I can find all kinds of ways to waste data. So, the analog waveform is converted to digital in the AD process and expressed as a series of numbers. The waveform is sampled at the sample rate, like 48000 times a second or 96000 times a second. Each of these samples is then stored as a number. The the precision of this number is determined by bit size. A 16 bit number is significantly smaller and therefore less precise than a 24 bit number. Right. So, in a nutshell. Moving from 16 bit to 24 bit, we have 8 extra bits per sample to represent the analog wave which is a massive gain. Not really. It gives you more dynamic range, which is often wasted anyway. 96 dB is an awful lot. Moving from 48khz to 96khz we simply double the number of bits we have. Which is why people will easily notice 16 bit vs 24 bit and less so 48khz vs 96khz. You might want to go back and read a good description of elementary sampling theory. I think Gabe references one in the FAQ. --scott -- "C'est un Nagra. C'est suisse, et tres, tres precis." |
#27
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 samplin
|
#28
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
|
#29
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
|
#30
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
Don Pearce wrote in message . ..
It is standard practice. The fact is that in any statistical test there will be results that stand out from the others. It is important to look at these results to see if they really are different, or merely the result of normal statistical clumping. Ah, I understand now. Thanks for the explanation. |
#31
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44sampling
mike rogers wrote:
Ok, has everyone lost the plot here or is it just me? So, what this test says is that basically music sounds the same regardless of whether or not it has everything above 21khz removed. So what the hell has this got to do with 24/48 or 24/96 recording? Nothing. If you take the study as gospel, it means that any difference between Fs = 48 kHz and Fs = 96 kHz or even Fs = 1 THz is inaudible, therefore those higher sample rates are a waste of money, disk space, and CPU time. If you don't agree with the study, then there may be value in those higher rates. So actually, it has everything to do with those sample rates' necessity in recording. |
#32
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
"mike rogers" wrote in message
This paper describes the test methodology and the results of a series of listening tests performed by researchers at NHK Science & Technical Research Laboratories, Tokyo, Japan. These tests compared the playback of recordings with and without audio signals above 21 KHz. Ok, has everyone lost the plot here or is it just me? So, what this test says is that basically music sounds the same regardless of whether or not it has everything above 21khz removed. So what the hell has this got to do with 24/48 or 24/96 recording? Nothing. Huh? 24/44 coding removes *everything* above 22 KHz. 24/96, 24/192 and 24/384 coding doesn't. They move the cut-off points to 48, 96, and 192 KHz respectively. Here is the simple computer programmers explanation of sampling rates. Analog sound is converted to digital. In order to do this it needs to be stored as bits. The more bits we can use to re-create the analog waveform, the better the sound. True only if the law of diminishing returns has been repealed. Furthermore there are two different and independent ways to use more bits to code audio signals. You can use more samples or you can use larger samples or both. All three options use more bits, but they have differing consequences. So, the analog waveform is converted to digital in the AD process and expressed as a series of numbers. The waveform is sampled at the sample rate, like 48000 times a second or 96000 times a second. Each of these samples is then stored as a number. The major benefit of this option is that the high frequency cutoff gets moved up in the frequency domain as the sample rate goes up. However it's not a sure thing that moving the high frequency cutoff up indefinitely provides improved sound quality. The law of diminishing returns has to start rearing its ugly head at some point. The precision of this number is determined by bit size. A 16 bit number is significantly smaller and therefore less precise than a 24 bit number. The major benefit of increased sample size is that the noise floor gets moved down in the amplitude domain as the samples get larger. However it's not a sure thing that moving the noise floor down indefinitely provides improved sound quality. The law of diminishing returns has to start rearing its ugly head at some point. This paper says nothing about this issue one way or the other. So, in a nutshell. Moving from 16 bit to 24 bit, we have 8 extra bits per sample to represent the analog wave which is a massive gain. The benefit is a reduced noise floor, or if you will higher resolution. However at some point the digital noise floor moves under the analog noise floor and further improvements are moot. Moving from 48khz to 96khz we simply double the number of bits we have. The benefit of an increased sample rate is an increased high frequency bandpass, or if you will an increased high frequency cutoff point. However the human ear is well known to lose accuracy and sensitivity above as little as 4 KHz. At some point so much accuracy and sensitivity is lost that further improvements the high frequency cutoff point become moot. This paper is about investigations into the benefits of moving the cutoff point well beyond 21 KHz. The investigations showed zero benefit for increasing the cutoff point beyond 21 KHz. This approximately corresponds to the real-world situation with a 44 KHz sample rate. The question the paper addresses is whether or not increasing the sample rate above 44 KHz (e.g. 96 KHz) has any audible benefits. It struggled diligently with the question and found no benefits to the major effect of increasing the sample rate substantially above 44 KHz. Which is why people will easily notice 16 bit vs. 24 bit and less so 48khz vs. 96khz. In actuality neither change is very easily noticed. If you have a DAW with 24/96 converters and a monitoring system you respect, you can investigate this with your own ears by downloading files from http://www.pbcabx.com/technical/sample_rates/index.htm and listening to them. Nothing to do with the peak human hearing frequency of 20khz as expressed in the start of this thread. I think I've explained why this is not a correct statement several different ways. |
#33
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
|
#34
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
|
#35
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 samplin
In article , "anthony.gosnell"
wrote: "Mike Rivers" wrote writes: At 11 Khz you have only 4 samples per cycle using 44.1Khz sampling frequency but nearly 9 samples per cycle using 96Khz. And this means? If you're suggesting that the 11 kHz sine wave will be more accurately reproduced from 9 samples than from four, you're wrong, provided that all other rules of sampling have been followed. Since when did music consist only of pure sine waves? Since forever. What do you think overtones are? A fundamental and a series of overtones can be broken down essentially into a bunch of sine waves. -- Jay Frigoletto Mastersuite Los Angeles promastering.com |
#36
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
|
#37
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44samplin
Jay - atldigi wrote: In article , "anthony.gosnell" wrote: "Mike Rivers" wrote writes: At 11 Khz you have only 4 samples per cycle using 44.1Khz sampling frequency but nearly 9 samples per cycle using 96Khz. And this means? If you're suggesting that the 11 kHz sine wave will be more accurately reproduced from 9 samples than from four, you're wrong, provided that all other rules of sampling have been followed. Since when did music consist only of pure sine waves? Since forever. What do you think overtones are? A fundamental and a series of overtones can be broken down essentially into a bunch of sine waves. What's too often forgotton is that a signal of finite length, like a song or a single drum hit, requires a whole bunch of them. An infinite number in fact. Conversely, anything that can be decomposed into a finite number of them must be infinitely long and repeat with a frequency equal to that of the lowest sin wave. Bob -- "Things should be described as simply as possible, but no simpler." A. Einstein |
#38
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
Don Pearce wrote in message . ..
Clearly being a simple programmer isn't sufficient. Moving from 16 to 24 bits improves matters if - and only if - the noise floor of the original analogue signal is below that of the 16 bit dither signal. The chances of that happening in any real recording are vanishingly close to zero. In virtually any scenario encountered in real life, 16 bits record just as high a quality as 24. There may well be special demo recordings that don't obey this rule of thumb. As regards sampling rate, the situation is perhaps not quite as clear. Certainly it is possible to find microphones that are pretty flat up to 20kHz, and have useful output above. Whether that makes any audible difference to a recording is debatable. Certainly recording above 20kHz will result in the capture of some pretty major untreated resonances from any microphone - manufacturers aren't quite as fussy about flatness up there. If you accept that 20kHz represents a useful upper limit to human hearing, and there is nothing significant above, then 44.1 is every bit as good as 48 or 96. 44.1 captures *everything* up to 20kHz with no exceptions. I agree with you regarding the noise floor on 16 bit recording. Your last paragraph is complete ******** though. I do understand sampling, having worked on software in this area. This is why you are wrong: Say we take a analog wave cycling at a fequency of 10khz or 10,000 times per second. We then sample that at 10khz. This means that for every 1 second of waveform time we take 10,000 samples to see what the amplitude of the waves is. From this information the computer can try to calculate what the wave actually "looked" like and reconstruct it in the DA process. But the analog world does not work in samples and there are actually an infinite number of possible sample points in a 1 second, 10khz wave. So, when we sample the same wave at 20khz, we now have a much more accurate representation of the orignal wave form as we have measured the amplitude in double the number of places so the wave recontruction is more faithful to the original. To truly represent an analog waveform, you would have to sample at infinite number of KHZ, which is obviously rediculous. At some point probably 96khz or a little bit above, no-one would be able to tell the difference. To prove my point, try this: Record an analog signal onto a PC, a higher pitched signal is better, at a low sample rate, like 8khz. Play the sample back through a spectral analyser and you will see frequencies above 8khz have been captured. According to what you are saying, it should be impossible to record audio frequencies higher than your sample rate. This is not true. True that the higher frequencies will not sound good as you will get a very poor representation of the higher frequency waveform, but they are still there. |
#39
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
The Sadie "ARTEMIS" is PC/Windows based.
Philip Perkins |
#40
|
|||
|
|||
Why 24/96 sampling isn't necessarily better-sounding than 24/44 sampling
|