I want to make (in CEP) a 20Hz to 20kHz bandpass filter to apply to my vinyl captures but I have 3 questions about filtering.
1. What is better suited for this purpose, an FFT filter or one of the scientific filters (e.g. Butterworth)?
2. Is there anything wrong with making it a brickwall filter or should the filter cutoff be gradual, e.g making the lower cutoff transition band start at 30Hz (100%) and finish at 20Hz (0%)?
3. Is it better to do a high pass filter then a low pass filter or a single bandpass filter?
Thanks in advance for any guidance.
I want to make (in CEP) a 20Hz to 20kHz bandpass filter to apply to my vinyl captures but I have 3 questions about filtering.
1. What is better suited for this purpose, an FFT filter or one of the scientific filters (e.g. Butterworth)?
The FFT filter is the best choice if you're looking for a one size fits all solution.
Sometimes you run into a filter that is hard to draw on the limited resolution graphic that the FFT provides. Then it the scientific filters can help. For narrow notches, such as removing hum, the DTMF filters really shine.
2. Is there anything wrong with making it a brickwall filter or should the filter cutoff be gradual, e.g making the lower cutoff transition band start at 30Hz (100%) and finish at 20Hz (0%)?
The general rule is do what sounds best.
3. Is it better to do a high pass filter then a low pass filter or a single bandpass filter?
Theroretically, you want to do as much as possible with as few processing steps. So, if you can get what you want with just one pass and a bandpass filter, then this is a good idea.
What bandlimiting is present in your (digital?) captures ? Perhaps the filtering in your ADC is already good enough.
WARNING - I'm not a DSP expert, so some of this might be wrong.
Why do you want/need to filter? I've never done that with my vinyl transfers. Mostly, I've worked on noise reduction, and sometimes I've used some EQ (high end boost) to fix-up old "dull sounding" recordings.
Any high frequencies (above Nyquist) should be filtered by the soundcard/ADC. (This may not be true for all soundcards, but any good soundcard should have an anti-aliasing filter.) And, if you're downsampling later, the downsampling algorithm/process will also include an anti-aliasing filter.
I'd guess the only reason for low frequency filtering would be subsonic rumble? I've never used a rumble filter, although a simple subsonic filter may have been built into my phono preamp. And, my noise-reduction processing may have removed some subsonic noise. I suppose it wouldn't hurt, but I'd probably set the cuttoff to something like 15Hz, just to preserve any 20Hz sounds... My speakers can't reproduce 20Hz, but I'd "feel better" about it.
I want to make (in CEP) a 20Hz to 20kHz bandpass filter...
Keep in mind that the specified cutoff frequency is the -3dB point. So, if you specify a "20 - 20kHz" filter, your signal will be down
exactly 3dB at 20Hz and at 20kHz (no matter what type of filter, or how steep it is).
1. What is better suited for this purpose, an FFT filter or one of the scientific filters (e.g. Butterworth)?
A Butterworth filter is probably a good choice. It has the flattest possible passband. This is a good choice whenever its more important
not to affect the passband than it is to filter-out stuff outside the passband. A different filter design might be appropriate whenever it's most important to kill some noise (or signal) outside the passband, and in order to accomplish that, you have to accept with some "ripple" in the passband.
You can do amazing things with an FFT filter, but if I understand how FFT filters work, they're "messy" and require lots of approximation and compromise... I think, just the FFT followed by inverse-FFT (without any filtering) is a lossy process... I'm pretty sure you can't get your exact original data back.
2. Is there anything wrong with making it a brickwall filter or should the filter cutoff be gradual, e.g making the lower cutoff transition band start at 30Hz (100%) and finish at 20Hz (0%)?
I think the only "cost' for a steep filter is the processing complexity and processing time (not an issue, since you are not working in real-time). But, there may be some phase/delay issues... I'm not sure.
3. Is it better to do a high pass filter then a low pass filter or a single bandpass filter?
I'm pretty sure the bandpass is made-up of combined high-pass and low-pass filters anyway... So, I wouldn't expect any difference.
There is intentionally nothing on an LP below 32Hz. Different turntable cartridge combinations have resonances that can cause skips with low frequencies so to avoid the issue, it is simply left out. I had an LP cut with 16-32 Hz classical organ and while it played fine on my Dual 721 with a Shure V-15 type V, other very expensive systems simply couldn't handle it.
A first or second order High pass with the corner at 30Hz would be reasonable for LPs.
Agreed on the HP / LP / bandpass issue. I'd do it in 1 pass also.
G²
Moderation: Removed unnecessary quotation of previous post.
> 1. What is better suited for this purpose, an FFT filter or one of the
> scientific filters (e.g. Butterworth)?
For my curiousity, what is a scientific filter? Is an FFT an unscientific filter? Are your FFT filters limited to powers of 2 in the number of samples involved?
As mentioned earlier, a high pass Butterworth filter at 30 Hz is a reasonable starting point for removing any low frequency rubbish that may be limiting the real signal. If your find you have quite a lot of it with your particular setup then a bit of experimenting with the cut-off point and slope might be worthwhile.
As mentioned earlier, I am not sure a low pass filter will do much for you since the A->D step should handle that efficiently.
Are you processing at a higher resolution/bit depth and then generating the target resolution/depth as a final step?
> 2. Is there anything wrong with making it a brickwall filter or should the
> filter cutoff be gradual, e.g making the lower cutoff transition band start at
> 30Hz (100%) and finish at 20Hz (0%)?
Brickwall filters will ring and this is often judged to be worse than a less efficient filter which rings less or, perhaps, not all.
> 3. Is it better to do a high pass filter then a low pass filter or a single
> bandpass filter?
In your case, it is unlikely to matter but 1 pass is probably going to be a bit quicker. However, for example, aggressive narrow band pass filters can have stability issues with some types of filters and so the option for more than 1 pass is usually necessary in a general setup.
For my curiousity, what is a scientific filter? Is an FFT an unscientific filter? Are your FFT filters limited to powers of 2 in the number of samples involved?
Scientific filter is just a name that the CEP authors gave to a class of filters. These filters are the classic canonical filters from the days of analog.
As mentioned earlier, a high pass Butterworth filter at 30 Hz is a reasonable starting point for removing any low frequency rubbish that may be limiting the real signal. If your find you have quite a lot of it with your particular setup then a bit of experimenting with the cut-off point and slope might be worthwhile.
Agreed that some kind of high pass filtering will often remove a lot of garbage.
When I filter the low end, I look to the recordng to be a guide as to how high to set the cutoff. If you do a FFT analysis of a whole track, the actual musical tones wiill cause distinct families of clearly visible spikes. Repetitive mechanical and electrical noises cause splkes at multiples of things like power line frequencies. Random mechanical and acoustical noise will cause relatively indistinct broad undulations.
CEP is very nice in that the filters have a preview filter. Click and drag a critical passage, select a filter, click the preview button, and adjust filter paramaters, wait for a second or two for them to become effective, and adjust until things sound right.
As mentioned earlier, I am not sure a low pass filter will do much for you since the A->D step should handle that efficiently.
Low pass filters are usually reserved for really noisy vinyl/
Are you processing at a higher resolution/bit depth and then generating the target resolution/depth as a final step?
IME, a waste of time unless you have a tic/pop algroithm that works better with really sharp-edged tics.
> 2. Is there anything wrong with making it a brickwall filter or should the
> filter cutoff be gradual, e.g making the lower cutoff transition band start at
> 30Hz (100%) and finish at 20Hz (0%)?
Brickwall filters will ring and this is often judged to be worse than a less efficient filter which rings less or, perhaps, not all.
The idea that brickwall filters necessarily ring is an urban myth. The cause of this myth is a misunderstanding that is caused by a misinterpretation of waveform pictures. Just because a picture of a square wave shows something that looks like a damped sine wave near leading and/or trailing edges, is not proof of ringing. This sort of thing can be caused by phase shifts.
> 3. Is it better to do a high pass filter then a low pass filter or a single
> bandpass filter?
In your case, it is unlikely to matter but 1 pass is probably going to be a bit quicker. However, for example, aggressive narrow band pass filters can have stability issues with some types of filters and so the option for more than 1 pass is usually necessary in a general setup.
Agreed.
The idea that brickwall filters necessarily ring is an urban myth. The cause of this myth is a misunderstanding that is caused by a misinterpretation of waveform pictures. Just because a picture of a square wave shows something that looks like a damped sine wave near leading and/or trailing edges, is not proof of ringing. This sort of thing can be caused by phase shifts.
Steady on - by definition, the sharper you make the cut off of a linear phase filter, the longer it will ring for. If the transition band is in the audible range, and there is original content in the vicinity of the transition band, you will hear the ringing.
If it's
not constrained to be linear phase, you can do funky things with the ringing - a usual trick is to minimise pre-ring, while allowing post-ring. It'll still ring though!
FWIW filtering out inaudible frequencies seems a bit of a waste of time, and can be very hit and miss - mainly because until you find the scenario where you
do need to remove them (e.g. you get much bigger/better speakers and can suddenly hear 20Hz rumble, or you apply some processing that's sensitive to these frequencies), you won't know whether your filtering did more good than harm.
Cheers,
David.
Oh, I've just noticed the thread title(!).
You don't need to apply a bandpass filter to put a 44.1kHz 16-bit wave file on a CD. Just stick it on there. CD is a 0Hz-22.05kHz medium, whatever the specs for most CD players say.
It won't cause any problems - unless there's a huge DC offset, in which case you might hear a click when skipping tracks.
Cheers,
David.
The idea that brickwall filters necessarily ring is an urban myth. The cause of this myth is a misunderstanding that is caused by a misinterpretation of waveform pictures. Just because a picture of a square wave shows something that looks like a damped sine wave near leading and/or trailing edges, is not proof of ringing. This sort of thing can be caused by phase shifts.
Steady on - by definition, the sharper you make the cut off of a linear phase filter, the longer it will ring for. If the transition band is in the audible range, and there is original content in the vicinity of the transition band, you will hear the ringing.
Not true. There's stuff happening, but there is no ringing.
You can do this for yourself in Cooledit/Audition.
(1) Pick a high sample rate like 24/196 so that any artifacts will be minimal.
(2) Generate a 10 Hz square wave at -20 dB peak, again so that any effects won't go off scale.
(3) Use the FFT filter to put in a brick wall filter (0 dB to -100 dB) low pass in at 6 KHz.
OK so you look at the resulting wave form and there appears to be some *ringing* around the leading edges.
I assert that yes there is what appears to be a damped sinusoid, but it isn't ringing.
If it is ringing, then there will be a peak in the spectral analysis, right?
So run a spectral analyis with 65 k points and show me the peak in response. What we expect to see is the odd harmonics of 10 Hz with amplitude inversely proportional to the frequency, and then they will be attenuated about 100 dB by the brick wall at 6 KHz.
If we look at the square wave, we see what appaears to be a damped rippling effect with a period of about one millisecond.
This is going to be like shooting ducks in a barrel, if it is ringing, there will be a peaking in response around 1 KHz. At one KHz the harmonics of the 10 KHz square wave are going to be coming thick and fast, so the ringing around 1 Khz will be *very* apparent, right?
It ain't there!
The apparent ringing is due to phase shifting of the harmonics of 10 Hz. To add up to be a square wave their phase and amplitude relationships of the harmonics must be very precisely like a square wave.
The brick wall filter messes that up. We don't get a pure square wave.
There is no ringing to hear. What we do hear is the straight-forward effect of the brickwall filter.
If it's not constrained to be linear phase, you can do funky things with the ringing - a usual trick is to minimise pre-ring, while allowing post-ring. It'll still ring though!
A messed up square wave does not necessarily prove the existence of any ringing.
I assert that yes there is what appears to be a damped sinusoid, but it isn't ringing. If it is ringing, then there will be a peak in the spectral analysis, right?
Absolutely not. If there were a peak, you would have
resonance. Resonance is not ringing. Ringing is temporal spreading - and a dampened sinusoid is quite obviously an instance of that. A long FFT is completely insensitive to that sort of thing.
FWIW, the preringing for foobar2000's FIR equalizer is ludicrously audible. I once (http://www.hydrogenaudio.org/forums/index.php?showtopic=26289)did a 5db peak followed by a 5db cut and could still ABX the ringing!
The apparent ringing is due to phase shifting of the harmonics of 10 Hz. To add up to be a square wave their phase and amplitude relationships of the harmonics must be very precisely like a square wave. The brick wall filter messes that up. We don't get a pure square wave. There is no ringing to hear. What we do hear is the straight-forward effect of the brickwall filter.
You're specifically referring to the Gibbs phenomenon here - phase shifts have nothing to do with this. The square wave just had its harmonics lopped off.
If the temporal smearing or ringing or dampenened sinusoids (or whatever you want to call them) exceed the temporal masking threshold, no matter what you call the phenomenon, it
will be audible. That should not be that hard to reproduce for filters exceeding 4000 samples for khz-range filtering. FIR filtering of rumble is usually a fool's errand, but I could imagine a suitably high-Q filter operating at 30hz to be plausibly audible with a suitable choice of listening equipment. I have no evidence to back that statement up though
I assert that yes there is what appears to be a damped sinusoid, but it isn't ringing. If it is ringing, then there will be a peak in the spectral analysis, right?
Absolutely not. If there were a peak, you would have resonance. Resonance is not ringing. Ringing is temporal spreading - and a dampened sinusoid is quite obviously an instance of that.
Now you are making a semantic argument, and established authories go against you:
http://en.wikipedia.org/wiki/Ringing (http://en.wikipedia.org/wiki/Ringing)
"In electrical circuits, ringing is an unwanted oscillation of a voltage or current."
When bells ring, it is due to resonances, but now you want to say that bells don't ring?
I'm not buying any! ;-)
A long FFT is completely insensitive to that sort of thing.
A long FFT of a complex wave, when processed by the inverse FFT transform, gives back a good reproduction of the original wave Ringing, and all. Making FFTs long does not make them inaccurate.
FWIW, the preringing for foobar2000's FIR equalizer is ludicrously audible. I once (http://www.hydrogenaudio.org/forums/index.php?showtopic=26289)did a 5db peak followed by a 5db cut and could still ABX the ringing!
Unfortunately I was unable to download the referenced files. They appear to have been deleted.
However, the presence of an audible difference does not guarantee that the difference exists for the reason hypothesized.
The apparent ringing is due to phase shifting of the harmonics of 10 Hz. To add up to be a square wave their phase and amplitude relationships of the harmonics must be very precisely like a square wave. The brick wall filter messes that up. We don't get a pure square wave. There is no ringing to hear. What we do hear is the straight-forward effect of the brickwall filter.
You're specifically referring to the Gibbs phenomenon here - phase shifts have nothing to do with this. The square wave just had its harmonics lopped off.
Right, and chopping off the higher harmonics disturbs the required precise phases and amplitudes. Not due to phase shift like I erroneously said, but rather due to massive screwing with the amplitudes. They're gone!
If the temporal smearing or ringing or dampenened sinusoids (or whatever you want to call them) exceed the temporal masking threshold, no matter what you call the phenomenon, it will be audible.
Are you aware of any modern brickwall filter in an >= 44KHz ADC or DAC that actually rings or smears outside of the multi-millisecond temporal masking interval?
That should not be that hard to reproduce for filters exceeding 4000 samples for khz-range filtering.
Evidence?
FIR filtering of rumble is usually a fool's errand, but I could imagine a suitably high-Q filter operating at 30hz to be plausibly audible with a suitable choice of listening equipment. I have no evidence to back that statement up though
A brick-wall rolloff below 30 Hz is going to be audible with many examples of musical program material simply because it is going to remove low frequency information that we hope is audible, or else it isn't worth getting rid of, or if it is music, preserving.
It's my belief that people who worry about equalizers having audible effects are missing the point of having equalizers! ;-)
Absolutely not. If there were a peak, you would have resonance. Resonance is not ringing. Ringing is temporal spreading - and a dampened sinusoid is quite obviously an instance of that.
Now you are making a semantic argument, and established authories go against you: <a href="http://en.wikipedia.org/wiki/Ringing" target="_blank">http://en.wikipedia.org/wiki/Ringing (http://en.wikipedia.org/wiki/Ringing)</a>
"In electrical circuits, ringing is an unwanted oscillation of a voltage or current." When bells ring, it is due to resonances, but now you want to say that bells don't ring? I'm not buying any! ;-)
OK, OK, maybe I did take a semantic turn there. I was speaking of resonance in terms of a second-order system, where a definite peak exists in the response. At the same time though, these damped sinusoids we are speaking of seem quite unwanted to me
A long FFT of a complex wave, when processed by the inverse FFT transform, gives back a good reproduction of the original wave Ringing, and all. Making FFTs long does not make them inaccurate.
And
you're the one complaining about semantics here? I'm talking about magnitude power spectra, not FFTs. Their graphical interpretation is entirely subject to debate. In particular, short-time power spectrum (shorter in time than the length of ringing) will clearly show frequency peaks in the transition band before/after the square wave transitions in your example; a long time power spectrum would not show it.
Unfortunately I was unable to download the referenced files. They appear to have been deleted. However, the presence of an audible difference does not guarantee that the difference exists for the reason hypothesized.
Quite true, but given knowledge of the implementation, it appears to be the most plausible explanation to me.
Are you aware of any modern brickwall filter in an >= 44KHz ADC or DAC that actually rings or smears outside of the multi-millisecond temporal masking interval?
No - I forgot to mention that I am aware of the research about phase shifts at brickwall frequencies etc being basically inaudible. I get that.
But the original discussion was about
rumble frequencies, right? The 10-30hz regime is a different ballgame entirely when it comes to audibility.
That should not be that hard to reproduce for filters exceeding 4000 samples for khz-range filtering.
Evidence?
<whips it out>Seriously, would you like for me to try to make a test of this? Perhaps implementing the same class of filter at two different numbers of taps, and demonstrating an ABX based solely on the ringing?
FIR filtering of rumble is usually a fool's errand, but I could imagine a suitably high-Q filter operating at 30hz to be plausibly audible with a suitable choice of listening equipment. I have no evidence to back that statement up though
A brick-wall rolloff below 30 Hz is going to be audible with many examples of musical program material simply because it is going to remove low frequency information that we hope is audible, or else it isn't worth getting rid of, or if it is music, preserving. It's my belief that people who worry about equalizers having audible effects are missing the point of having equalizers! ;-)
I tended to agree, until I figuratively threw the foobar2000 eq against a wall a few years ago. There are real disadvantages to excessively high-Q filters.
For the record, when I derumble, I use a 6th order Butterworth at 25hz for L+R, and an 8th order Butterworth at 35hz for L-R.
I've attached a .zip archive containing 5 files
ringing_original.wav - a series of impulses to filter.
(I could have used almost any broadband signal, but impulses make the effect really obvious.)
ringing_4kHz_LPF_brickwall.wav - the effect of filtering the signal with a brick wall filter. I used the FFT filter in CEP - the corresponding .gif file shows the settings.
ringing_4kHz_LPF_gentle.wav - the effect of filtering the signal with a more gentle filter. Again, the corresponding .gif file shows the settings.
I didn't take much care with this, but it proves my point: brick wall filtering rings, and if the transition band is within the audible range, this ringing is audible.
Take a listen for yourself - I can clearly hear the ringing at 4kHz in the brick wall filtered version. And yes, it does sound a bit like a bell!
The version with the gentle filter avoids this (though it's not a very careful attempt to do so - just a quick application of the "spline curves" feature in CEP).
I wrote a response to some of the points raise, but my PC crashed. No bad thing. There's no real answer to some of the points, because they're so...!
E.g. the suggestion that ringing should be examined via a 65k FFT! Even the worst temporal smearing any audio codec has ever created would be invisible under such examination - you don't examine temporal effects using long FFTs!
As for "what is ringing" - "that's not ringing, it's Gibbs" - ?!?!?!!!!!
Cheers,
David.
P.S. Please, no more talk of phase shifts - the CEP FFT filter isn't doing any. It's linear phase.
I've attached a .zip archive containing 5 files
ringing_original.wav - a series of impulses to filter.
(I could have used almost any broadband signal, but impulses make the effect really obvious.)
ringing_4kHz_LPF_brickwall.wav - the effect of filtering the signal with a brick wall filter. I used the FFT filter in CEP - the corresponding .gif file shows the settings.
ringing_4kHz_LPF_gentle.wav - the effect of filtering the signal with a more gentle filter. Again, the corresponding .gif file shows the settings.
I didn't take much care with this, but it proves my point: brick wall filtering rings, and if the transition band is within the audible range, this ringing is audible.
Take a listen for yourself - I can clearly hear the ringing at 4kHz in the brick wall filtered version. And yes, it does sound a bit like a bell!
I don't hear a bell sound at all -- I hear the two filtered versions both sounding considerably more 'muffled' than the original. And both sounding a little different from each other too.
(using foobar2k v 9.5.3, output at 16 bit Direct Sound , no dither, to onboard SigmaTel Audio on a Dell PC, into Koss TD61 headphones)
I don't hear a bell sound at all -- I hear the two filtered versions both sounding considerably more 'muffled' than the original. And both sounding a little different from each other too.
They are 'muffled' just because they are lowpassed. But "brickwall" version also contains some hiss (around 4 kHz).
I've attached a .zip archive containing 5 files
ringing_original.wav - a series of impulses to filter.
(I could have used almost any broadband signal, but impulses make the effect really obvious.)
ringing_4kHz_LPF_brickwall.wav - the effect of filtering the signal with a brick wall filter. I used the FFT filter in CEP - the corresponding .gif file shows the settings.
ringing_4kHz_LPF_gentle.wav - the effect of filtering the signal with a more gentle filter. Again, the corresponding .gif file shows the settings.
I wonder how many controls have been involved in any listening tests comparing these files, as they are supplied with 3 significantly different amplitudes, the closes still about 0.5 dB apart.
I normalized them, checked average RMS to make sure that they didn't different much in that direction, doubled their length by cutting and pasting, and then ABXd them. I could easily hear a difference and immediately scored 16/16, but the difference was not night and day.
I also checked the frequency response of the two filtered ways and found that they varied by upwards of 6 dB in the transition band. The transition band for the gentle filter is about 200 Hz wide for a center frequency of 4 KHz, which may be enough to explain much of why they sound different. They are very different filters and may sound different simply because of the obvious difference.
Using impulses for a test like this seems like an extreme worst case. Music with impusive percussion would be more like a reasonable worst case.
I wonder how many controls have been involved in any listening tests comparing these files, as they are supplied with 3 significantly different amplitudes, the closes still about 0.5 dB apart.
That's surely going to happen with impulses (e.g. approximations to a Dirac delta function).
The energy of a delta function is evenly distributed in frequency (just like white noise). If you filter out everything above 4 kHz you're going to lose a lot of energy (leaving 4000/22050 = 18% of the original total energy if it's sampled at 44100 Hz), and lose a good deal of peak amplitude, having spread the impulse temporally. That's what it's supposed to do. The two filters probably pass a similar proportion of the total energy, but one has a gentle cut-off that should cause little temporal spread and the other is a brick wall that should create a lot of temporal spreading.
The main thing we're trying to do is listen for any ringing and temporal spreading effects, not really to ABX and calibrate volume levels.
I wonder how many controls have been involved in any listening tests comparing these files, as they are supplied with 3 significantly different amplitudes, the closes still about 0.5 dB apart.
That's surely going to happen with impulses (e.g. approximations to a Dirac delta function).
Of course.
The main thing we're trying to do is listen for any ringing and temporal spreading effects, not really to ABX and calibrate volume levels.
I'm very surprised to hear people on Hydorgen Audio dismissing the idea of introducing scientific controls into their listening tests.
I guess this is just a game, and without any serious intent?
I don't hear a bell sound at all -- I hear the two filtered versions both sounding considerably more 'muffled' than the original. And both sounding a little different from each other too.
They are 'muffled' just because they are lowpassed.
Ah, also the sound of me being dumb.
But "brickwall" version also contains some hiss (around 4 kHz).
I heard that as a tiny bit of extra 'detail' compared to the gentle filter, rather than as a bell.
The two filtered versions were easy to ABX, whether 'raw' or level-matched with Replaygain.
(With replyagain's boosting, --~ +25 dB -- the 'detail' became a whistling)
foo_abx 1.3.3 report
foobar2000 v0.9.5.3
2008/12/04 18:37:48
File A: ringing_4kHz_LPF_brickwall.flac
File B: ringing_4kHz_LPF_gentle.flac
without replaygain
18:42:27 : Test started.
18:42:41 : 01/01 50.0%
18:43:07 : 02/02 25.0%
18:43:12 : 03/03 12.5%
18:43:16 : 04/04 6.3%
18:43:19 : 05/05 3.1%
18:43:23 : 06/06 1.6%
18:43:26 : 07/07 0.8%
18:43:41 : 07/08 3.5%
18:43:49 : 08/09 2.0%
18:43:56 : 09/10 1.1%
18:44:00 : 10/11 0.6%
18:44:09 : 11/12 0.3%
18:44:19 : 12/13 0.2%
18:44:23 : 13/14 0.1%
18:44:26 : 14/15 0.0%
18:44:30 : 15/16 0.0%
18:44:36 : 16/17 0.0%
18:44:39 : Test finished.
----------
Total: 16/17 (0.0%)
with replaygain:
18:37:48 : Test started.
18:38:25 : 01/01 50.0%
18:38:32 : 02/02 25.0%
18:38:37 : 03/03 12.5%
18:38:43 : 04/04 6.3%
18:38:46 : 05/05 3.1%
18:38:50 : 06/06 1.6%
18:38:53 : 07/07 0.8%
18:38:57 : 08/08 0.4%
18:39:01 : 09/09 0.2%
18:39:07 : 10/10 0.1%
18:39:11 : 11/11 0.0%
18:39:14 : 12/12 0.0%
18:39:18 : 13/13 0.0%
18:39:22 : 14/14 0.0%
18:39:27 : 15/15 0.0%
18:39:30 : 16/16 0.0%
18:39:33 : 17/17 0.0%
18:39:42 : Test finished.
----------
Total: 17/17 (0.0%)
I heard that as a tiny bit of extra 'detail' compared to the gentle filter, rather than as a bell.
The two filtered versions were easy to ABX, whether 'raw' or level-matched with Replaygain.
(With replyagain's boosting, --~ +25 dB -- the 'detail' became a whistling)
Well, it's that whistling I called 'hiss'... It is really simple to ABX; and this effect of steep filter is obvious at spectrograms (left is the signal after gentle filter, right - after brickwall)
(http://img254.imageshack.us/img254/3162/84941483mw4.png)
I don't hear a bell sound at all -- I hear the two filtered versions both sounding considerably more 'muffled' than the original. And both sounding a little different from each other too.
They are 'muffled' just because they are lowpassed.
But "brickwall" version also contains some hiss (around 4 kHz).
I heard that as a tiny bit of extra 'detail' compared to the gentle filter, rather than as a bell.
The two filtered versions were easy to ABX, whether 'raw' or level-matched with Replaygain.
(With replyagain's boosting, --~ +25 dB -- the 'detail' became a whistling)
A whistling?
I hear the difference between my two level-matched samples as the brickwall version having a little more snap, or detail.
The two samples still don't have the same frequency response +/- 0.1 dB. Therefore, of course they are going to sound different.
Occam's razor suggests that the obvious frequency response difference is the cause of the audible difference, not the differences in ringing.
If we're going to attribute the difference to ringing and not frequency response, then the frequency response of the two samples needs to be matched, with the only difference being the amount of ringing.
The two samples still don't have the same frequency response +/- 0.1 dB. Therefore, of course they are going to sound different.
Are you seriously saying that the human ear can detect amplitude errors within the passband as small as 0.2dB, or am I misinterpreting this? I thought it was closer to 2dB. Feel free to poke me in the eye with a sharp stick if I'm being an eejit.
Cheers, Slipstreem.
A whistling?
I hear the difference between my two level-matched samples as the brickwall version having a little more snap, or detail.
I'd expect that's because your level matching didn't boost them by ~+25dB, as replaygain does.
The two samples still don't have the same frequency response +/- 0.1 dB. Therefore, of course they are going to sound different.
Are you seriously saying that the human ear can detect amplitude errors within the passband as small as 0.2dB, or am I misinterpreting this? I thought it was closer to 2dB. Feel free to poke me in the eye with a sharp stick if I'm being an eejit.
Cheers, Slipstreem.
Nope, it's 0.2dB , in the part of the passband that humans are most sensitive to (midrange). Here's a web tutorial that appears to corroborate the figure
http://www.avatar.com.au/courses/PPofM/loud/Loud1.html (http://www.avatar.com.au/courses/PPofM/loud/Loud1.html)
The minimum change in SPL required to give a detectable change in the loudness sensation (JND in sound level) is roughly constant and of the order of 0.2 - 0.4 dB in the musically relevant range of pitch and loudness.
I suspect the lower figure (0.2 dB) is from experiments using pure tones, but don't have a reference for that.
Aren't the samples supposed to have differing peaks? The total energy in each pulse should be very nearly equal. Normalizing them is an incorrect procedure here - in a very real sense, they're already normalized.
The two samples still don't have the same frequency response +/- 0.1 dB. Therefore, of course they are going to sound different.
Are you seriously saying that the human ear can detect amplitude errors within the passband as small as 0.2dB, or am I misinterpreting this? I thought it was closer to 2dB. Feel free to poke me in the eye with a
In this particular case there were differences of as much as 5 dB in the upper part of the pass band, between the brickwall and so-called gentle filter. I don't know if they are audible or not, but they are big enough so that their presence creates an unanswered question.
In general, we set a +/- 0. 1 dB tolerance, because some variations as small as 0.5 dB are actually pretty easy to hear under some conditions. It's a overkill number, but not very far into overkill.
2 dB is perceptually *huge* compared to the type of differences that are routinely audible in ABX tests.
There are a lot of conditions on what predicts the potential audibility of a FR differnce. They were presented in detail by means of a chart in the 1978 JAES article by Clark that introduced ABX. For example, narrow dips may be hearder to hear than narrow peaks, and narrow variations may be harder to hear than varaitions over wide ranges. Variations near 4 KHz may be easier to hear than variations near 20 Hz or 20 KHz.
When I listen to the three files successively in fb2k I hear:
ti ti ti ti ti ti (ringing_original.wav) - sharp transient with generally high pitch
pp pp pp pp pp pp (ringing_4kHz_LPF_gentle.wav) - a series of duller transients with slightly lower perceived pitch.
then the same duller sound,
pp pp pp pp pp pp (ringing_4kHz_LPF_brickwall.wav)
superimposed with similtaneous bright metallic ringing, similar to a teaspoon hitting a mug or more like a metal cold-chisel being hit by a hammer, the resonance damped quickly by a gloved hand.
The difference is night and day - no need for ABX.
I'm glad that some HA regulars were around to answer the points raised by Arny's posts. I was away for the weekend, so couldn't respond.
This worried me...
I'm very surprised to hear people on Hydorgen Audio dismissing the idea of introducing scientific controls into their listening tests.
I guess this is just a game, and without any serious intent?
... in fact this is quite sad for me. For various reasons, I'd come to regard you Arny as a world authority on blind testing (despite never having corresponded with you). However, your every contribution to this thread is so wide of the mark, and so correctly rebutted by HA regulars, that I guess HA (or its members) just became the
de facto world authority.
If you feel any of the points you made haven't been adequately answered, please say which ones you feel have been overlooked, and I shall attempt to address them.
(Anyone else is free to jump in - you're all doing a great job! )
Cheers,
David.
P.S. If you're near London, you might be interested to hear tomorrow's AES lecture / interview (it's a "Christmas special"!), which I believe may include discussion of the choice of minimum phase filters instead of linear phase filters for anti-image filtering. I assume this is to reduce pre-ringing, though do not understand how this could be audible at 22kHz!
http://www.aes.org/sections/uk/meetings/index.html#1208 (http://www.aes.org/sections/uk/meetings/index.html#1208)
I'll probably be there. This does not mean I'm turning into a subjectivist
...which I believe may include discussion of the choice of minimum phase filters instead of linear phase filters for anti-image filtering. I assume this is to reduce pre-ringing, though do not understand how this could be audible at 22kHz!
That could be an interesting evening. In the ongoing discussions about hi-res (and vinyl?) audio benefits the filters remain an important element. If you do go, could you ask some suggestions for killer samples that can demonstrate filter effects (even in a DBT )? I've been thinking about setting up a test but it's a tricky one. If the effect can't be demonstrated in a sighted test, like you did with your 4kHz samples, it's probably a waste of time. Could you please ask as well about speakers vs headphone listening ? Thanks and enjoy the evening.
ps: it would be nice if the "supporting material" (hopefully audio samples) is freely available.
The difference is night and day - no need for ABX.
Very very many of the golden ears say the exact same thing about their listening tests relating to green pens, magic cables, etc.
So, I guess we're supposed to tell the difference between what the golden ears say, and what you say because... ????
People who "get it" just do the listening test with the relevant controls and report statisitically significant results.
It's pretty obvious who around here gets it, and who doesn't. And, going to AES lectures doesn't seem to help. It's this way in most technologies. Lip service versus actual change in behavior.
I see a line-by-line critique is needed. I'll work on it.
btw, I didn't report "going to AES lectures" to impress, or to imply I had some credentials. For one thing, I've been to a couple in the last decade, and for another, anyone who has ever attended one would not claim that attendance is anything to boast about!
Cheers,
David.
Firstly, I believe in double blind testing. Any genuinely audible difference should be able to be proven as such in a statistically verifiable double blind test.
There may be practical issues (e.g. DBT of speakers is a challenge), but most issues can be overcome with sufficient effort.
Now, let's talk about the current test.
Firstly, should we ABX when "the difference is obvious"? Well, on HA we generally let people off from ABXing "obvious" things, but where there's doubt, people should be happy to ABX for the avoidance of doubt. TOS 8! krabapple has already ABXed, so that's that issue answered.
Secondly, is it appropriate to ABX this? I would say no - most people are claiming an audible difference, and you yourself are not doubting an audible difference. It's the
nature of the audible difference that's in doubt, and ABX cannot help here. Blind A/B testing may be useful.
Thirdly, should we level match? Yes, for a fair comparison, we should. However, your methods of doing so are misguided. Where the audible differences may be subtle, and/or where the mathematical change to the signal may be significant, blind application of peak or RMS normalisation may
increase the perceived level difference between signals, rather than remove it. ReplayGain would be an answer if it was perfect, but it is not. It might be the "least bad" solution in some circumstances, but not here.
The best solution available is to manage all processes that introduce a level difference, and mitigate them - or to design the experiment to avoid introducing level differences entirely. In the current example, the energy within the flat part of the passband is identical in all three examples. For this type of signal, that's about as good as it gets. There are corrections you could apply for the non-flat part of the pass band, but that would break this experiment by introducing an audible difference into the flat part of the pass band.
Fourthly(!), you demand the frequency response be matched to +/-0.2dB. If we were ABXing two things that should sound the same and/or be the same, this would be relevant - but in this case we are comparing a brick wall filter (which you say does not ring, and I say does) with a non-brick wall filter (which I say rings far less). These have
different time and frequency domain responses. This is the nature of the experiment.
If I were to take a brick wall filter, and try to conduct this test whilst staying with +/-0.2dB of the brick will filter, then the
other filter would also be a brick wall filter - I would be comparing two brick wall filters! Not very useful, I'm sure you'll agree.
Fifthly, you said...
Very very many of the golden ears say the exact same thing about their listening tests relating to green pens, magic cables, etc.
So, I guess we're supposed to tell the difference between what the golden ears say, and what you say because... ????
Let's get one thing very clear: green pens, magic cables etc create no
measurable, objective difference. Whereas the difference we're examining here is visible in the waveform plot, and in a spectrogram. Also, if you look at, and measure, the audio signal, and compare it to the nearest relevant psychoacoustic data, you'll see that the ringing is predicted as being audible.*
Maybe
you can't hear it. It's well known that absolute thresholds vary between individuals, and that
masked thresholds vary even between individuals with identical absolute thresholds. I would speculate that sensitivity to
temporal masking varies even more widely between individuals than sensitivity to spectral masking.
However, it's been ABXed, and multiple posters have now reported hearing the same thing. Added to the fact that it's visible on a spectrogram, this demolishes the idea that it's imagined.
I think you've spent too long around subjectivist audiophiles. Your knee jerk "magic cable" response is incorrect and unjustified.
Your original assertion - that brick wall filters do not ring - belies such a fundamental misunderstanding of filtering and hearing that I expect this thread to run for a while yet. But let's see if we can't resolve our differences over the ABXing of these files first.
Cheers,
David.
* That's despite the fact that the nearest data uses a burst of white noise, while the current test uses a tiny click! Even the white noise can't drag the masking thresholds sufficiently high in a pre-masking experiment in order to mask the ringing measurably present in a brick wall filtered signal.
Pre masking data, white noise masking tone, here:
Gehr, S. E.; and Sommers, M. S. (1999).
Age difference in backward masking.
Journal of the Acoustical Society of America, vol. 106, no. 5, Nov., pp. 2793-2799.
I went to the lecture last night, and am happy to post a report, but I'd rather get the above issues resolved first.
Cheers,
David.
Fifthly, you said...Very very many of the golden ears say the exact same thing about their listening tests relating to green pens, magic cables, etc.
Let's get one thing very clear: green pens, magic cables etc create no measurable, objective difference.
Actually, magic cables generally do have measurable differences. Speaker cables of any signficiant length are even a little like amplifiers, in that they all measure differently, but most if not all of the measurable differences are too small or of the wrong kind to matter.
Many interconnects are just a little microphonic. The stuff we can reliably measure today can be just crazy!
Green pens have a number of measurable effects on CDs, but they rarely if ever show up in the audio domain.
The point is that trying to defend not doing an ABX test on the grounds that there was a measurable effect just doesn't work out in real life.
The goal is to reduce subjectivity and judgement calls as much as possible. Of course, we still end up making judgement calls.
Your original assertion - that brick wall filters do not ring - belies such a fundamental misunderstanding of filtering and hearing that I expect this thread to run for a while yet. But let's see if we can't resolve our differences over the ABXing of these files first.
One of your real serious problems is that you make a lot of presumptions, and you personalize way too much.
You presume that my assertion is an indication of defficiencies in my basic education. In fact my education ain't all that bad. I've probably known about, worked with, and had some fairly useful understanding of the Gibbs phenominon for maybe 45 years.
Nahh, what happened is that I've been thinking about this for several weeks just lately, taking a sort of a blank paper approach.
I've been told for decades that the brick wall filters in 44 Khz ADCs and DACs clearly and audibly corrupted sound quality, but I know for a fact that a well-done pair of 44 KHz ADCs and DACs are sonically transparent in demanding real-world tests.
Of course they all appear to ring to some degree. What's your explanation for all the apprarent ringing and yet all the ABX tests with null results?
Last time I challenged established thinking about digital like this, I came up with the concept of self-dither, which is now rather widely recognized. My main opponent in that debate was very, very abusive. He got to be wrong. ;-)
I've been told for decades that the brick wall filters in 44 Khz ADCs and DACs clearly and audibly corrupted sound quality, but I know for a fact that a well-done pair of 44 KHz ADCs and DACs are sonically transparent in demanding real-world tests.
Just for kicks, you ought to look at some of the half-band filters used in some of the common sigma-delta convertors out there. I wouldn't use the term "brick wall", though. This may be a separate discussion, though, of actual implimentation compared to how it should be done.
Of course they all appear to ring to some degree. What's your explanation for all the apprarent ringing and yet all the ABX tests with null results?
Ringing is nothing more than the impulse response of any sharp filter. FIR, IIR, whatever, if it's a sharp (in terms of small transition band) filter, it's going to "ring".
Is it audible?
It is POSSIBLE (but unproven as far as I know, barring some results being established here, perhaps) that exactly the wrong kind of ringing, perhaps from a not-well-designed FIR filter could lead to perceptable pre-echo under some very, very special conditions...
Please notice all of the qualifications in that statement.
Arny,
Let's get back on the same page. I've made no claim that anti-alias filters are audible.
I challenged your statement that brick wall filters do not ring.
I said...
the sharper you make the cut off of a linear phase filter, the longer it will ring for. If the transition band is in the audible range, and there is original content in the vicinity of the transition band, you will hear the ringing.
To back this up, I provided some samples with a brick wall filter at 4kHz.
http://www.hydrogenaudio.org/forums/index....st&p=602827 (http://www.hydrogenaudio.org/forums/index.php?s=&showtopic=67619&view=findpost&p=602827)
I think the ringing (or whatever you want to call it) is clearly audible. "Dynamic" seems to hear the same. You seem to hear something different. What do other people think/hear?
I've been told for decades that the brick wall filters in 44 Khz ADCs and DACs clearly and audibly corrupted sound quality, but I know for a fact that a well-done pair of 44 KHz ADCs and DACs are sonically transparent in demanding real-world tests.
Of course they all appear to ring to some degree. What's your explanation for all the apparent ringing and yet all the ABX tests with null results?
The ringing is at 22kHz. I can't hear 22kHz.
That's why I provided samples with a cut-off frequency of 4kHz.
We're confusing two completely different issues in this thread - audibility of
ultrasonic ringing, and audibility of
ringing within the audible band. I initially thought you were only talking about the former, but then you provided an example to prove that ringing at 6kHz was inaudible - I don't accept this at all.
No more personal comments. I don't have any "problem" with you - I was just holding you to a far higher standard than anyone else due to your background, so made far harsher criticisms of things I didn't agree with. I apologise.
Cheers,
David.
I think the ringing (or whatever you want to call it) is clearly audible. "Dynamic" seems to hear the same. You seem to hear something different. What do other people think/hear?
I agree the ringing at 4k is clearly audible. I tried increasing the filter frequency and it becomes less audible---ABXing would be needed at higher frequencies.
Back at 4k, I also tried sliding the phase response around, and using an intermediate phase response, managed to achieve a reduction, but not elimination, of the audible ringing. Minimum phase still sounds pretty bad (when the source signal is clicks).
-bandpass
I think the ringing (or whatever you want to call it) is clearly audible. "Dynamic" seems to hear the same. You seem to hear something different. What do other people think/hear?
I also hear a pitched noise around 4kHz in the brickwall version, which is hardly audible in the gentle version.
The difference is night and day - no need for ABX.
Very very many of the golden ears say the exact same thing about their listening tests relating to green pens, magic cables, etc.
So, I guess we're supposed to tell the difference between what the golden ears say, and what you say because... ????
People who "get it" just do the listening test with the relevant controls and report statisitically significant results.
It's pretty obvious who around here gets it, and who doesn't. And, going to AES lectures doesn't seem to help. It's this way in most technologies. Lip service versus actual change in behavior.
2bdecided "gets it". He's most definitely a 'white hat' on audio matters, as are most of the regular posters here.
And HA is most definitely not Usenet -- suspicions and accusations appropriate for, say, the rec.audio groups, aren't applicable here.
I think the ringing (or whatever you want to call it) is clearly audible. "Dynamic" seems to hear the same. You seem to hear something different. What do other people think/hear?
I agree the ringing at 4k is clearly audible. I tried increasing the filter frequency and it becomes less audible---ABXing would be needed at higher frequencies.
I too heard a 'brighter' or slightly metallic presentation in the gentler filter, compared to the steeper one, in my ABX comparisons. There was also a sensation that the gently filtered sample was pitched slightly higher than the steeply-filtered one; these perceptions were how I was able to ABX them, without level-matching.
The difference is night and day - no need for ABX.
Very very many of the golden ears say the exact same thing about their listening tests relating to green pens, magic cables, etc.
So, I guess we're supposed to tell the difference between what the golden ears say, and what you say because... ????
People who "get it" just do the listening test with the relevant controls and report statisitically significant results.
It's pretty obvious who around here gets it, and who doesn't. And, going to AES lectures doesn't seem to help. It's this way in most technologies. Lip service versus actual change in behavior.
2bdecided "gets it".
I'm sticking to what I said. People who "get it" just do the listening test with the relevant controls and report statisitically significant results.
I reported the first ABX test results from this particularly comparison, showing that I practice what I preach.
He's most definitely a 'white hat' on audio matters, as are most of the regular posters here.
To some degree. He seems to be naive about measurable differences in audio cables, for example.
And HA is most definitely not Usenet -- suspicions and accusations appropriate for, say, the rec.audio groups, aren't applicable here.
I dunno, I've had more than enough suspicions and false accusations thrown at me so far. :-(
Not nearly as bad as RAO (little is!), but probably worse than RAP.
I'm sticking to what I said. People who "get it" just do the listening test with the relevant controls and report statisitically significant results.
Fair enough, but we might need your help.
Two stimuli are presented that are known to be different. Their frequency spectra differ by more than 0.1 dB in the audible range. IIRC you stated earlier in this thread that an ABX test is not valid when the spectra are not within a 0.1 dB margin. How can this test be made valid in your opinion ?
For that matter, I'm still not convinced that Arny's normalization process (and therefore his ABX test) was ever valid to begin with. It's my understanding that the passband spectrum was exactly the same between all the samples, so the normalization process actually denormalized them.
I'm sticking to what I said. People who "get it" just do the listening test with the relevant controls and report statisitically significant results.
Fair enough, but we might need your help.
Two stimuli are presented that are known to be different. Their frequency spectra differ by more than 0.1 dB in the audible range. IIRC you stated earlier in this thread that an ABX test is not valid when the spectra are not within a 0.1 dB margin. How can this test be made valid in your opinion ?
ABC/hr ?
For that matter, I'm still not convinced that Arny's normalization process (and therefore his ABX test) was ever valid to begin with. It's my understanding that the passband spectrum was exactly the same between all the samples, so the normalization process actually denormalized them.
Seems like a moot point since it is a matter of fact that the two signals vary by several dB in the passband.
So far nobody has matched the levels of these signals in such a way that they were indistingushable in an ABX test. Closer analysis suggests that could well be mission impossible. Perhaps this is not the test that will resolve the question at hand?
The experiment that would resolve this question would process the input signal twice, each time producing a signal whose spectral energy content, etc. matched the other processed signal very closely.
The only difference between the two signals would be that one processed signal had the waveform abberations that are typical of the Gibbs Phenominon, and the other did not.
It's been suggested that there may exist some minimum phase filter that is less intrusive than other non-minimum phase filters, but provides the same output spectral energy content. The phase shifting of the two signals can be expected to be different.
It's been suggested that there may exist some minimum phase filter that is less intrusive than other non-minimum phase filters, but provides the same output spectral energy content. The phase shifting of the two signals can be expected to be different.
Are you using a "linear phase" (constant delay) filter as your non-minimum-phase filter? Are these FIR's, if so how long?
Given a set of coefficients, one can create (for any zero pair that is not on the unit circle) another FIR that is more closely minimum-phase than a standard symmetric FIR quite trivially, or in fact create a maximum-phase filter as well.
Still done via FIR, but resulting in non-symmetry.
A short example:
Columns 1 through 9
0.0047 0.0213 0.0045 -0.0388 -0.0581 0.0227 0.1970 0.3467 0.3467
Columns 10 through 16
0.1970 0.0227 -0.0581 -0.0388 0.0045 0.0213 0.0047
versus
cc =
Columns 1 through 10
0.0153 0.1036 0.2420 0.3467 0.3107 0.1353 -0.0528 -0.1211 -0.0607 0.0264
Columns 11 through 16
0.0518 0.0205 -0.0095 -0.0128 0.0033 0.0015
These two filters have identical (magnitude) frequency response to the level of rounding in the printout. I'm not feeling energetic enough to print them to double precision right now.
The only question in this process is "is Matlab's rooting algorithm able to do the filter roots right". If the answer is yes, invert all zeros with magnitude over 1, and you get as close to minimum phase as you can get.
The experiment that would resolve this question would process the input signal twice, each time producing a signal whose spectral energy content, etc. matched the other processed signal very closely.
The only difference between the two signals would be that one processed signal had the waveform abberations that are typical of the Gibbs Phenominon, and the other did not.
With a linear phase filter, this is impossible. The ringing is proportional to the steepness of the transition band. If the "spectral energy content" matches, then, for linear phase, the amount of ringing will match. It will be two copies of the same filter - there are no other variables.
It's been suggested that there may exist some minimum phase filter that is less intrusive than other non-minimum phase filters, but provides the same output spectral energy content. The phase shifting of the two signals can be expected to be different.
Indeed so, but that would allow you to compare two (or more) brick wall filters with the ringing shifted about due to phase differences. This is all very interesting, and I could guess with some certainty at the audible result, but this isn't the exact point I was making.
It's easy enough to do, but requires MATLAB or similar, not Cool Edit Pro - unless you want to try one of its IIR filters, one of which might be minimum phase, and time reverse the result, which hence might be maximum phase.
Any takers?
Cheers,
David.
It's easy enough to do, but requires MATLAB or similar, not Cool Edit Pro - unless you want to try one of its IIR filters, one of which might be minimum phase, and time reverse the result, which hence might be maximum phase.
Any takers?
Here you go: phases.zip (http://www.shada.plus.com/phases.zip)
000 = minimum phase, ..., 050 = linear phase, ..., 100 = maximum phase
-bandpass
Thanks bandpass.
The difference is night and day.
Minimum phase sounds like someone flicking/tapping a slightly damped wine glass, as you'd expect. Sharp start, then decay. The other two have a soft start.
For the doubters, I confirm that the magnitude response is identical, though you need a very long FFT to check it.
How did you generate these? (Not that I doubt - just interested).
Cheers,
David.
How did you generate these? (Not that I doubt - just interested).
Well, it might sound like doing it the hard way, but it was something I wanted to do anyway: I extricated the filter generator from the resampler in the latest SoX source code and ran it on its own.
An interesting question is at what phase response is the ringing least audible---for me it seems to be around 030.
-bandpass
Certainly stacking them up in fb2k and playing through them, the click seems most click-like with least ringing around file 030. What do the numbers correspond to?
If you look at the waveform, 030 has less post-ring than 000, but no pre-ring. In fact it's the shortest filter with no pre-ring at all, which probably explains the preference.
I always "thought" minimum phase was the shortest filter, so maybe I don't understand.
Cheers,
David.
Certainly stacking them up in fb2k and playing through them, the click seems most click-like with least ringing around file 030. What do the numbers correspond to?
The numbers are percentage phase response; so 000 is minimum phase and 100 is maximun phase. You'll need to use a dB amplitude scale with the impulse to see what's going on---linear amplitude scale won't cut it.
E.g. 030 does have pre-ring, but you won't see it with linear amplitude.
-bandpass
I think there are 2 different claims to prove about this ringing issue.
(1) Whether brickwall filtering causes ringing or not.
(2) If there is ringing, whether this ringing is audible or not.
As for (1), let's restrict just to linear phase filters for simplicity. Linear phase filters are filters that have no phase distortion, or in other words, filters that don't modify phase is at all (but for a simple constant time delay to the signal). A linear phase brickwall filter is no other thing than the sinc function (http://en.wikipedia.org/wiki/Sinc_function). As you can see, the fourier transform, or in other words, the frequency domain view of this function is the rectangular function (http://en.wikipedia.org/wiki/Rectangular_function), this is, a function where there the result is 1 up to a frequency and is 0 from this frequency to infinite, or the maximum sampled frequency of this system. This is obviously the same than a brickwall filter.
As you can see, the time-domain view of this funcion has what is commonly called pre and post ringing, a fading-in and fading-out sinusoid of the same frequency than the filter cutoff.
All this can be simulated and verified in CEP or Adobe Audition creating an impulse signal, brickwall filtering it with a FFT filter, and looking al the resultant signal. In frequency domain you can see the ringing using the spectral view, or using a short FFT analysis (128 point for example) in the temporal proximities of the filtered impulse.
So, from my point of view, point (1) is proved.
Now let's go for (2), is it audible? Well, the question for me is if there is something that prevents it from being audible. The answer is that there are two things that can prevent it:
The first one is the frequency of the ringing. If it is outside of the audible range, then it is inaudible. So brickwall filters over 20 KHz will not be audible, which is the case of cd-audio antialiasing and reconstruction filters.
The second thing is the length of the ringing. Due to a propery of out ear called temporal masking, a faint signal may get masked by another louder signal very close in time, this is what is called temporal masking. Faint signals before a loud signal need to be very close to this loud signal to get masked, in other words, our ear is not very prone to backwards temporal masking. But faint signals after the loud signal may be more away from the loud signal and still get masked, in other words, our ear is more prone to forward temporal masking.
So, if the ringing due to the brickwall filtering is at an audible frequency but is very short in time, it will be very probably masked due to this temporal masking property of our hearing. If the ringing is not very short but not very long, the pre-ringing will be audible but the post-ringing not. And if it is very long, both pre and post ringing may get audible.
What does the lenght of the ringing depends on? Just on the length of the filter, in other words, the number of samples used to compute it, which is the same as the order of the filter. A 65K-samples filter will have 32K samples of pre-ringing and 32K samples of post ringing. A long filter is sharper in frequency domain, but has longer ringing. A short filter is less sharp in frequency domain, but has shorter ringing. A long "softened" brickwall filter will have shorter ringing too.
So, is it audible? Depends on the cutoff of the brickwwall, the lenght of the filter and the sampling frequency. At 44.1 KHz and a brickwall cutting at an audible frequency, A 16K filter has 0.2 s of pre-ringing, which is easily audible over an impulse. A 512 samples filter has 6 ms of pre-ringing, which is much less audible, but probably still audible. A 128 samples filter is less than 2 ms, I think it is not audible. A 256 maybe yes, maybe not.
I think the difference between the 16K and the 512-samples filter is very audible. See the attached file with examples of the results of these filters at a 4KHz cutoff over an impulse signal. I had no time to ABX, but the difference is obvious. The files are time-aligned and level matched at 0.01 dB or less, so no processing is needed. Both the level match of signals in the baseband and frequency response of filters can be checked doing a dual-channel FFT of the stereo file. So (2) is proved too.
Finally, note that if the signal to be filtered has no content at the cutoff frequency, there will be no ringing to talk of. And if the signal is not impulsive, the ringing will be nearly nonexistent compared with the non-impulsive signal.
Sorry for any typos, at the end I was on a hurry when writing this.
I agree with all of that. Especially the "why wouldn't it be audible" approach - I've already referenced the temporal masking data on this.
But you've dragged us back to the "harder" question of the audibility of ringing in two filters with non-identical frequency response - i.e. both linear phase, one "gentle". I have no problem with this, but this is where I started two pages ago, and was criticised, so I'd retreated to the even "safer" position of the audibility of ringing in two filters with identical frequency response.
Still, in either case, I can't believe anyone can plausibly claim the ringing from a brick wall filter is not audible when the transition band is within the audible range, and there's content within the original signal to exercise the ringing. I hope Arny will be gracious enough to retract his former claim and accept that this is the case.
Cheers,
David.
Yes, I restricted to linear phase in order to focus in the two different claims over this ringing issue, and the importance of the filter length over the audibility. It's hard to believe the audible differences between the long and short filter are not due to ringing, because the audible difference is that one file rings and the other does not (or not as much). Also, your files were properly level matched in the pass band, Arny's normalization over the whole band probably destroyed this level matching.
As you say, going the case of exact same frequency response and just changing phase of the filter (and associated location of ringing) makes even more evident that the audible differences are due to these different locations of ringing, so the conclusion is that this ringing is audible.
Still, in either case, I can't believe anyone can plausibly claim the ringing from a brick wall filter is not audible when the transition band is within the audible range, and there's content within the original signal to exercise the ringing. I hope Arny will be gracious enough to retract his former claim and accept that this is the case.
I'm very impressed and convinced by the data. There's no question about the level matching or frequency response matching. The samples are totally matched in the pass band, vary only enough to convince me that they are really different in the transition band, and differ measurably but inconsequentially in the rejection band. Of course the temporal data is definitive and obvious.
Listening test data for at least the 2 extreme cases is equally clear and definative. ABXing to any reasonble level of statistical difference is trivially easy (even using my ancient ears diminished by a bad head cold).
Furthermore, the 5% and 100% samples clearly and unmistakably sound like they are whistling. The 100% sample seems to be whistling around 4 KHz and the 5% sample a little less, a little lower. The linear phase filter seems to minimze this effect.
In terms of temporal dispersal, the linear phase filter seems to have reduced temporal dispersal, even as compared to the original wave.
Since the 50% is well-qualified as a brick wall filter, there is no obvious proof that brick wall filters are necessarily evil.
One of the less obvious characteristics of filter is the degree to which it causes the data to have increased peak amplitude in either direction. Again, the linear phase filter seems to be the winner.
I hear and see a lot of justification for linear phase filtering.
Now let's go for (2), is it audible? Well, the question for me is if there is something that prevents it from being audible. The answer is that there are two things that can prevent it:
The first one is the frequency of the ringing. If it is outside of the audible range, then it is inaudible. So brickwall filters over 20 KHz will not be audible, which is the case of cd-audio antialiasing and reconstruction filters.
I just thought this should be emphasized.
In terms of temporal dispersal, the linear phase filter seems to have reduced temporal dispersal, even as compared to the original wave.
Since the 50% is well-qualified as a brick wall filter, there is no obvious proof that brick wall filters are necessarily evil.
So you don't hear any chirping with the linear phase filter?
And you think the 4kHz brick wall linear phase filtered click has "reduced temporal dispersal, even as compared to the original wave"?
Cheers,
David.
The numbers are percentage phase response; so 000 is minimum phase and 100 is maximun phase. You'll need to use a dB amplitude scale with the impulse to see what's going on---linear amplitude scale won't cut it.
Here's a graph (time x-axis, amplitude dB y-axis) of the phases 0-50% (minimum to linear phase). Phases 55-100% are not shown but look like phases 45-0% mirrored about the y-axis.
Due to differences between forward and backward temporal masking (http://en.wikipedia.org/wiki/Temporal_masking), linear phase should not sound the best to most people. (But I agree that this does not apply to filters @ >20kHz).
-bandpass
(http://i39.tinypic.com/r2on4n.gif)
You'll need to use a dB amplitude scale with the impulse to see what's going on---linear amplitude scale won't cut it.
Missed that first time around - thanks for re-stating it!
IIRC there's a comment by Brian Moore that temporal pre-masking decreases greatly with training, making the pre/post difference even more stark. Something like 3ms vs 20ms for a masker/maskee combination, but that's from memory - I've lost the paper.
Cheers,
David.
In terms of temporal dispersal, the linear phase filter seems to have reduced temporal dispersal, even as compared to the original wave.
Since the 50% is well-qualified as a brick wall filter, there is no obvious proof that brick wall filters are necessarily evil.
So you don't hear any chirping with the linear phase filter?
Vastly less, as compared to the 5% and 100% samples. Comparison with the original wave is invalid because it has a vastly different frequency response.
And you think the 4kHz brick wall linear phase filtered click has "reduced temporal dispersal, even as compared to the original wave"?
Again, comparison with the original wave is obviously invalid, as it has vastly different frequency response.
I think we may need to review the purpose of the exercise. We are trying to filter out HF noise as effectively as possible with minimal effect on SQ.
Observing that no filter affects sound less than a potentially effective filter doesn't seem to help very much.
So you don't hear any chirping with the linear phase filter?
Vastly less, as compared to the 5% and 100% samples.
OK, that's fine, no one can argue with that.
I think we may need to review the purpose of the exercise. We are trying to filter out HF noise as effectively as possible with minimal effect on SQ.
We were, but we've moved on a long way from the original question. We're in the
audible range now, for one thing.
So let me turn this around.
My hypothesis is that a brick wall filter rings, and that this ringing can be audible and objectionable. A more gentle filter will ring substantially less, to the point where the ringing is substantially inaudible.
The spectral and temporal responses of the gentle filter are different from those of the brick wall filter, but whereas the change in spectral response is small and rarely of any practical importance (I don't say whether it's audible or not), the change in temporal response is dramatic, clearly audible, and solves an objectionable problem with brick wall filtering.
How would you prove / disprove this in a way that you would find convincing?Suitable samples are at the end of KikeG's post:
http://www.hydrogenaudio.org/forums/index....st&p=604669 (http://www.hydrogenaudio.org/forums/index.php?s=&showtopic=67619&view=findpost&p=604669)
Cheers,
David.
[quote name='2Bdecided' date='Dec 16 2008, 06:20' post='604891']
[quote name='Arnold B. Krueger' post='604877' date='Dec 16 2008, 11:14']
[quote name='2Bdecided' post='604792' date='Dec 15 2008, 12:36']So you don't hear any chirping with the linear phase filter?[/quote]
Vastly less, as compared to the 5% and 100% samples.[/quote]
OK, that's fine, no one can argue with that.
[/quote]
I think we may need to review the purpose of the exercise. We are trying to filter out HF noise as effectively as possible with minimal effect on SQ.[/quote]
We were, but we've moved on a long way from the original question. We're in the
audible range now, for one thing.[/quote]
We always were in the audible range, when we're talking effective filters in the middle of the frequency range of most sensitive human hearing.
OTOH, we always have to keep our eyes on our masking curves.
The impulse is composed of every frequency that isn't being filtered out, so conceivably it is masking concurrent sounds in the same frequency range around the time it has peak amplitude. That potentially includes the ringing.
The impulse is also a very loud sound at the time its amplitude peaks, so it is masking sounds that are happening at about the same time. That potentially includes the ringing.
I suspect that if we can appropriately control the duration, amplitude and shape of the envelope of the ringing, we might mask it all.
My hypothesis is that a brick wall filter rings, and that this ringing can be audible and objectionable.
You do understand that the filter need not be brick wall in order to ring. As a matter of fact, our test signal rings pretty vigorously with no filtering at all. Also, I applied a 20 Hz impulse to a 4th order filter, and it rang as well. Just less, and more clearly under a presumed temporal masking curve.
A more gentle filter will ring substantially less, to the point where the ringing is substantially inaudible.
But, more noise gets through.
The spectral and temporal responses of the gentle filter are different from those of the brick wall filter,
Major side effect - remove less noise for a given effect on the music.
but whereas the change in spectral response is small and rarely of any practical importance (I don't say whether it's audible or not),
I believe that it was audible for a more gentle filter that was proposed earlier in the thread.
the change in temporal response is dramatic, clearly audible, and solves an objectionable problem with brick wall filtering.
But of course, less filternig, less noticable.
How would you prove / disprove this in a way that you would find convincing?
First I think we need to agree on what we are expecting the filter to do in terms of its first order effect.
The ringing, etc., is a second order or higher effect.
Suitable samples are at the end of KikeG's post:
http://www.hydrogenaudio.org/forums/index....st&p=604669 (http://www.hydrogenaudio.org/forums/index.php?s=&showtopic=67619&view=findpost&p=604669)
Suitable for what?
Those samples only focus on the secondary effects, and show nothing about any balance between primary and secondary effects.
As long as low slope, no slope filtering has no cost associated with it, it will always be the means of choice. After all, the best way to do nothing is to simply do nothing. ;-)
BTW this post is a mess of quotes, but I've stopped trying to please this conference software. I can balance quotes on other HTML forums, no sweat. There seem to be some hidden agendas.
At least on Usenet, we've got tools that handle quoting automagically.
BTW this post is a mess of quotes, but I've stopped trying to please this conference software. I can balance quotes on other HTML forums, no sweat. There seem to be some hidden agendas.
At the top of your post, you opened three quotes, then closed five. At least that is what is shown.
Cheers,
David.
I suspect that if we can appropriately control the duration, amplitude and shape of the envelope of the ringing, we might mask it all.
But when constrained to a linear phase brick wall filter, we have no independent control of any of those things - they are all pre-defined. A linear phase brick wall filter is a sinc function. That is what it is. The only variables are the cut off frequency, and where you decide to truncate the time-domain waveform (since an ideal sinc function is infinite).
KikeG has already covered making it "less brick wall" by truncating/windowing the response. This is the entire point - doing this makes the filter more gentle in its frequency response and (obviously) shorter in its temporal response.
You do understand that the filter need not be brick wall in order to ring.
The vast majority of filters ring. What's that got to do with the price of fish? The problem is whether the ringing is audible, and can be reduced so that it is not. Linear phase brick wall filters create audible ringing unless you make them less brick wall like!
A more gentle filter will ring substantially less, to the point where the ringing is substantially inaudible.
But, more noise gets through.
Ah, so we come to the question of what the filter is
supposed to be doing, and whether a sharp cut off is more important than the avoidance of ringing.
That depends on the application. The point is whether the brick wall filter introduces audible ringing. You stated that it did not. So while you suggest I'm assigning a "zero cost" to the gentle cut off, you are assigning a "zero cost" to the ringing.
In truth, both have a cost - judging which is preferable is a matter for the designer, and application specific.
However, that's was not your original point at all: you implied that ringing is a "zero cost" problem because you believe there is no ringing, and if there is, it isn't audible.
I'm not asking you how you would prove to your satisfaction "what is the best filter for this job" or even "what is this job" - I'm asking
how would you prove or disprove that the ringing from brick wall filters is audible.
Cheers,
David.
BTW this post is a mess of quotes, but I've stopped trying to please this conference software. I can balance quotes on other HTML forums, no sweat. There seem to be some hidden agendas.
At the top of your post, you opened three quotes, then closed five. At least that is what is shown.
Cheers,
David.
Not just that, there appears to be a ten-quote limit (brickwall filter? ) per post --I've run into it before, with complex replies.
It's not an 'agenda', Arny, though it might indicate that post could use trimming.
Here's a graph (time x-axis, amplitude dB y-axis) of the phases 0-50% (minimum to linear phase). Phases 55-100% are not shown but look like phases 45-0% mirrored about the y-axis.
The iZotope RX application can display time and spectrum at the same time. Perhaps it helps to see the differences even better.
I've picked the phases from 0-100% from left to right in 10% increments.
(http://www.galaxyclassics.com/public/4kHzLowPassFilters.jpg)
IIRC there's a comment by Brian Moore that temporal pre-masking decreases greatly with training, making the pre/post difference even more stark. Something like 3ms vs 20ms for a masker/maskee combination, but that's from memory - I've lost the paper.
Yes, I remember reading that somewhere too. Bit reluctant to try it myself though--don't want to end up hearing more artefacts! I daresay the actual masking periods are very signal dependent, filtered clicks being one of the most provocative cases.
The iZotope RX application can display time and spectrum at the same time. Perhaps it helps to see the differences even better.
I've picked the phases from 0-100% from left to right in 10% increments.
Nice display. For the hell of it, I've redone my time-domain display above as an animated gif.
Does anyone fancy extracting the nuggets from this thread to a wiki entry?
-bandpass
(http://www.galaxyclassics.com/public/4kHzLowPassFilters.jpg)
That's a beautiful graph, but only the centre one is linear phase - not all of them, as first words of the text imply.
Cheers,
David.
only the centre one is linear phase - not all of them, as first words of the text imply.
Thanks for pointing that out. It has been corrected.
Ah, so we come to the question of what the filter is supposed to be doing, and whether a sharp cut off is more important than the avoidance of ringing.
That depends on the application. The point is whether the brick wall filter introduces audible ringing. You stated that it did not. So while you suggest I'm assigning a "zero cost" to the gentle cut off, you are assigning a "zero cost" to the ringing.
No, I'm saying that there are two costs that need to be balanced - balance the cost of ringing to the cost of having more audible noise.
In truth, both have a cost - judging which is preferable is a matter for the designer, and application specific.
Check the title of the thread - the application is "Mastering Captured Vinyl For CD".
However, that's was not your original point at all: you implied that ringing is a "zero cost" problem because you believe there is no ringing, and if there is, it isn't audible.
Apparently you haven't kept up with my recent recantation of that position.
What I've learned so far is that it would probably be a good thing if we had filtering tools for vinyl mastering to CD that were based on linear phase filters, or better yet filters that were tipped a bit from linear phase towards minimum phase so that the ringing receives maximum temporal masking.
It further appears that said tools might benefit from variable slope features so that a subjective trade-off could be made to balance noise and ringing.
I know of no such tools on the market today, but it appears that they would be feasible to develop.
Voila: a product opportunity.
Apparently you haven't kept up with my recent recantation of that position.
Well, it is a little difficult to keep up with your position on anything...
In terms of temporal dispersal, the linear phase filter seems to have reduced temporal dispersal, even as compared to the original wave.
Comparison with the original wave is invalid because it has a vastly different frequency response.
I think we may need to review the purpose of the exercise. We are trying to filter out HF noise as effectively as possible with minimal effect on SQ.
We were, but we've moved on a long way from the original question. We're in the audible range now, for one thing.
We always were in the audible range, when we're talking effective filters in the middle of the frequency range of most sensitive human hearing.
Check the title of the thread - the application is "Mastering Captured Vinyl For CD".
Quitting now, before the forum quote bug strikes!
What I've learned so far...
I was hoping what you've learnt so far is that brick wall filters ring, the ringing can be audible and objectionable, and it can be dramatically reduced by making the filters less brick-wall like.
You know, all the stuff that some of us were saying from page one
...is that it would probably be a good thing if we had filtering tools for vinyl mastering to CD that were based on linear phase filters, or better yet filters that were tipped a bit from linear phase towards minimum phase so that the ringing receives maximum temporal masking.
It further appears that said tools might benefit from variable slope features so that a subjective trade-off could be made to balance noise and ringing.
I know of no such tools on the market today, but it appears that they would be feasible to develop.
Voila: a product opportunity.
Making the filters more gentle, rather than fiddling with the phase response, is the dramatically preferably solution
unless the transition from objectionable noise to wanted signal is extremely localised spectrally.
This isn't the case with Vinyl: brick wall filters are pointless here. It's not even the case with 78s or tapes - the noise is broadband, while the signal level reduces (somewhat gently) at high frequencies. If any filtering is appropriate, it's gentle filtering. Brick wall filtering is not only pointless in this instance, but the associated audible ringing will be triggered by any transient (or worse still, impulsive) content in the recording. Do you think a recording from vinyl or shellac might contain any impulsive signal elements?
As for a product opportunity: Yes.
However, Cool Edit Pro has had the FFT filter for over a decade - you can dial in as sharp or as gentle a filter as you wish. You cannot change the phase, but since the ringing is inaudible with a reasonably gentle filter, it would be of no great benefit there.
For the people trying to filter out spectrally localised noise (e.g. 15.625kHz squeak, 50/60Hz hum), shifting the phase would be beneficial. Though there can be better ways of defeating both of those issues, depending on the exact interference.
I daren't mention that one product opportunity appears to be in brick wall filters for CD upsampling. The top end of the market appears to be moving / have moved from brick wall to gentle, and from linear phase towards minimum phase. The difference "should" be inaudible, yet in A/B (not ABX) tests, the preference for these inaudible filters matches the preference for audible filters demonstrated in this thread - closer to minimum than linear phase; more gentle than brick wall.
I'm not claiming anything. Just pointing out the "co-incidence".
Cheers,
David.
The difference "should" be inaudible, yet in A/B (not ABX) tests, the preference for these inaudible filters matches the preference for audible filters demonstrated in this thread - closer to minimum than linear phase; more gentle than brick wall.
So, do we have the courage (after 3 pages of warming up) to design a test to find out if >20kHz brick wall filter artifacts can be audible ?
I'd be surprised if tests like this have not been done before, so perhaps we can learn from them.
That should be as easy as constructing a lowpass filter with a 1hz transition band, and then finding somebody with the barest inkling of 20khz hearing, right?
That can't be that hard to put together. (It's considerably more difficult to actually relate it to practical filters.)
The difference "should" be inaudible, yet in A/B (not ABX) tests, the preference for these inaudible filters matches the preference for audible filters demonstrated in this thread - closer to minimum than linear phase; more gentle than brick wall.
A/B tests, you mean sighted evaluations?
We can't rely on just gentle slope filters because we need notches to get rid of hum and other similar interfering signals.
So, do we have the courage (after 3 pages of warming up) to design a test to find out if >20kHz brick wall filter artifacts can be audible ?
44 Khz sample rate audio is endemic. It is almost totally based on > 20 KHz brick wall filters. Most good implmentations are sonically transparent, that is they defy detection in ABX tests.
I'd be surprised if tests like this have not been done before, so perhaps we can learn from them.
This is the earliest test of > 20 KHz brick wall filtering that I know of - late 1970s, or very early 1980s:
http://www.provide.net/~djcarlst/abx_digi.htm (http://www.provide.net/~djcarlst/abx_digi.htm)
If CEP (mentioned in 1st post) means Cool Edit Pro and, if CEP supports VST plug-ins and if you plan to record the filtered output then, there's a plug-in called RubberFilter (http://www.savioursofsoul.de/Christian/?page_id=64) made by Christian-W. Budde. Rubberfilter provides up to 64th order Butterworth filters to ‘rubber’ out frequencies.
If CEP does not support recording the FX output then, at least Coscos Reaper supports this feature.
If a VST implementation is OK for you then, there's also Butterworth and Cebyshev HP/LP modules for Delphi and SynthEdit available (same site). SynthEdit (http://www.synthedit.com/) is easy to use VST creation environment ... it took less than half of an hour for me to build a "twin" HP filter type of 1-64th order Butterworth/1-32th order Cebyshev.
C-W. Budde's SE modules (http://www.savioursofsoul.de/Christian/?page_id=442) and Delphi modules (http://sourceforge.net/projects/delphiasiovst/).
Schematic (http://img126.imageshack.us/img126/6/seschema2tx4.jpg) I used for twin HPF (similiar schematic works w/ LPF as well). Here's the twin HPF (http://www.zshare.net/download/52716431946566bc/) I prepared for testing purpose.
Here are some results from measures I did:
- Roll-of speed comparison (http://img84.imageshack.us/img84/5196/hocomparisoniz1.jpg)
- Large test (http://jiiteepee.fortunecity.com/tests/HP/simple_HP_filter_test.html) using Christian-W. Budde's VST Plugin Analyzer (http://www.savioursofsoul.de/Christian/?page_id=106)
Depending on audio interface, is it necessary to add a LPF?
Juha
So, do we have the courage (after 3 pages of warming up) to design a test to find out if >20kHz brick wall filter artifacts can be audible ?
Nah, we've got non-linear distortion to get through yet
There's some - not very scientific - evidence that steep anti-aliasing filters degrade certain subtle aspects of audio quality, although ringing takes place in the ultrasound range: "Effects in High Sample Rate Audio Material" (http://www.audiofast.pl/pdfs/effects.pdf).
Hi Alex,
it seems that these were sighted comparisons. Although an interesting attempt, and correlating with many professional engineers' opinions, it's not an acceptable proof in this forum's sense. The differences attributed to the filters could (QED) completely disappear in a double blind test.
Do you have any plans to evaluate filter effects (e.g. ringing) for src.infinitewave.ca ?
Since most DACs are linear phase and therefore suffer from ringing effects themselves, what would be a good setup to test for audible effect ? (my guess: upsampling to as high as the DAC allows)
There's some - not very scientific - evidence that steep anti-aliasing filters degrade certain subtle aspects of audio quality, although ringing takes place in the ultrasound range: "Effects in High Sample Rate Audio Material" (http://www.audiofast.pl/pdfs/effects.pdf).
IMO, the key text from that piece which I first downloaded from the DCS web site is:
"Recording engineers, and many musicians – particularly in the classical area – are becoming aware
that material recorded and edited using these higher sample rates has some attractive qualities.
Current theory on how human hearing works has so far been unable to explain the basis for these
qualities, but they are none the less easy to demonstrate."
The key phrase is "...they are none the less easy to demonstrate". Anybody who has spent as much time and money as I have spent out of my own personal pocket (and the pockets of my friends), in sincere and sell-supported attempts to demonstrate the possible benefits of increased bandpass on sonic accuracy, can't avoid having a bit of an emotional reaction.
I translate the quoted paragraph above as reading "Long-established science says that sample rates higher than 44.1 KHz have no audible benefts, but we're from DCS and our business plan is based on selling overpriced DACs for fun and profit, so we're going to throw stones at established science and see if enough people buy into our little anti-scientific song and dance to make us rich."
At this point, the audio industry had its little fling with SACD and DVD-A, and it is now pretty clear that at least some of the time, you can't fool enough people to make a good mainstream business out of fooling them.
Hi Alex,
it seems that these were sighted comparisons. Although an interesting attempt, and correlating with many professional engineers' opinions, it's not an acceptable proof in this forum's sense. The differences attributed to the filters could (QED) completely disappear in a double blind test.
Do you have any plans to evaluate filter effects (e.g. ringing) for src.infinitewave.ca ?
Since most DACs are linear phase and therefore suffer from ringing effects themselves, what would be a good setup to test for audible effect ? (my guess: upsampling to as high as the DAC allows)
I've participated in a number of attempts to show the non-transparency of various brick wall filters operating with a 22 KHz bandpass.
The first time was a few years before the CD came on the market, probably 1979 or 80. Our source material was live studio sound and 15 ips half track masters. The filters in question were part of a digital delay that Ampex was marketing to people who had automated cutting lathes. I suspect the brick wall filters were implemented in the analog domain, and were closest to being minimum phase.
Later on I did a bunch of demos for the now-defunct PCABX web site. The origional source material was recorded @24/96 using B&K 4006 measurement mics. The brick wall filters were implemented with CoolEdit's FFT filter.
IMO, the interesting question is not whether 22 KHz brick wall filters are sonically transparent, but rather how low can you push brick wall filters before they start being noticable.
IME with most music (one exception being KikeG's little ca. 16 KHz resonator) a 16 KHz brick wall low pass can be pretty innocent. It seems like a lot of people building perceptual coders tend to agree that < 22 Khz is OK.
If CEP (mentioned in 1st post) means Cool Edit Pro and if CEP supports VST plug-ins
CEP does not support VST plugs.
Audition, CEP's sequel, supports VST plugs in its current version.
Audacity can be a useful platform for running VST plugs.
I agree with you. I never had the chance to hear the differences myself. However many people claim that they do hear differences of different SRC algorithms, for example, depending on the filter type used (again all above 20 kHz).
I agree with you. I never had the chance to hear the differences myself. However many people claim that they do hear differences of different SRC algorithms, for example, depending on the filter type used (again all above 20 kHz).
I question some of the means that people use to hear differences related to different SRC algorithms. One approach is to take some regular music in say the 24 bit domain, attenuate it by 40 dB, downsample it to 16 bits, amplify it by 40 dB, and then present the results for comparison.
Do you think that is representative of how people actually use SRC software?
SRC converts sampling rate, not the bit depth. However your test looks good for testing of bit depth reduction. One important point is that listening levels are adjusted so that the dithering noise is kept inaudible or barely audible in this test - just as it is in most real situations.
SRC converts sampling rate, not the bit depth.
Right, but SRC is a generic term, used for software that changes both wordlength and sample rate.
However your test looks good for testing of bit depth reduction.
Really, do you seriously think that is how people use SRC software in day-to-day applications?
One important point is that listening levels are adjusted so that the dithering noise is kept inaudible or barely audible in this test - just as it is in most real situations.
Real situations where dither noise is audible? I can't think of one situation where the background noise on a CD I was listening to was related to dither. Room noise always dominates in real world recordings, doesn't it?
I can't think of one situation where the background noise on a CD I was listening to was related to dither. Room noise always dominates in real world recordings, doesn't it?
Obviously, it depends on your listening levels and the dynamic range of the recording.
I can't think of one situation where the background noise on a CD I was listening to was related to dither. Room noise always dominates in real world recordings, doesn't it?
Obviously, it depends on your listening levels and the dynamic range of the recording.
If you're routinely listening at levels where dither noise is audible, you're probably damaging your hearing.
Did I say I am? I suggested the opposite: dithering noise should be inaudible.
Did I say I am? I suggested the opposite: dithering noise should be inaudible.
My 'you' was meant universally, but if you took it personally, my apologies.
So, do we have the courage (after 3 pages of warming up) to design a test to find out if >20kHz brick wall filter artifacts can be audible ?
I'd be surprised if tests like this have not been done before, so perhaps we can learn from them.
Maybe "bandpass" would be happy to re-run the filter generation he did on page 2 of this thread, but with a sampling rate of 96kHz, and a low pass filter cut-off frequency of 20kHz - giving us minimum phase, maximum phase, linear phase (050, as previously labelled), and (for the sake of it) 030 and 070.
The resulting impulses can be listened to as-is, or convolved with 96kHz material to implement the appropriate filtering. Note convolution normally involves time domain reversal, so minimum and maximum phase will swap - unless the particular convolution implementation explicitly re-reverses the filter coefficients.
Cheers,
David.
So, do we have the courage (after 3 pages of warming up) to design a test to find out if >20kHz brick wall filter artifacts can be audible ?
I'd be surprised if tests like this have not been done before, so perhaps we can learn from them.
Maybe "bandpass" would be happy to re-run the filter generation he did on page 2 of this thread, but with a sampling rate of 96kHz, and a low pass filter cut-off frequency of 20kHz - giving us minimum phase, maximum phase, linear phase (050, as previously labelled), and (for the sake of it) 030 and 070.
The resulting impulses can be listened to as-is, or convolved with 96kHz material to implement the appropriate filtering. Note convolution normally involves time domain reversal, so minimum and maximum phase will swap - unless the particular convolution implementation explicitly re-reverses the filter coefficients.
Cheers,
David.
Okay, here they are (http://www.shada.plus.com/phases-20k.zip) with transition bandwidth = 0.16% nyquist (as before).
-bandpass
Thank you bandpass.
I've started a new thread and posted some 20kHz brick wall filtered samples to ABX (or not!):
http://www.hydrogenaudio.org/forums/index....showtopic=68524 (http://www.hydrogenaudio.org/forums/index.php?showtopic=68524)
Cheers,
David.
If you're near London, you might be interested to hear tomorrow's AES lecture / interview (it's a "Christmas special"!), which I believe may include discussion of the choice of minimum phase filters instead of linear phase filters for anti-image filtering. I assume this is to reduce pre-ringing, though do not understand how this could be audible at 22kHz!
http://www.aes.org/sections/uk/meetings/index.html#1208 (http://www.aes.org/sections/uk/meetings/index.html#1208)
I'll probably be there. This does not mean I'm turning into a subjectivist
There is an mp3 of the entire Bob Stuart lecture, downloadable from here:
http://www.aes.org/sections/uk/meetings/a0812.html (http://www.aes.org/sections/uk/meetings/a0812.html)
Most lectures have mp3s available...
http://www.aes.org/sections/uk/meetings/past.html (http://www.aes.org/sections/uk/meetings/past.html)
The quality of the recordings is terrible. This isn't a reflection on the AES, but the venue the meetings are held in.
There are some interesting points, which I'm sure many members of HA would argue with.
Cheers,
David.
Here are the (selective) notes I made on the night. (re-listening to the mp3 might reveal they're inaccurate, but I don't have two hours to spare!)
I didn't note anything that I was already familiar with. Everything below (unless marked) is my recollection of his opinion, and not my own.
It's a report of a lecture, not a statement of fact. No TOS strikes please!
2008-12-09 AES UK Bob Stuart Meridian
Active Loudspeakers. He doesn't like correcting response errors narrower than the auditory filter bandwidth. It's amazing what you can hear through distortion - e.g. you might have 10% speaker distortion, 0.01% amplifier distortion, and you can still hear the error in the 22nd bit of a filter.
Room correction. He think this is safe below ~200-300Hz, but treacherous above that. For one thing, it smears the time domain. For another, he believes speakers should be in the same room as the listener. We don't equalise real people talking in a room to correct for the room!
DVD-A. He wanted to give something back to the industry that was worthwhile, and better than CD. Before DVD-A he was involved with Xtrabits (http://www.aes.org/e-lib/browse.cfm?elib=7964), the work for which revealed the required noise floor. There's a paper: Coding for High Res audio systems (http://www.aes.org/e-lib/browse.cfm?elib=12986), AES, March 2004.
He discussed the required noise floor using several frequency/amplitude plots showing the threshold of hearing (lower curve), the maximum peak replay level (straight line at the top, he set this to 114dB SPL), and the audible distortion or noise of various systems.
16-bit TPDF dither just exceeds the threshold of hearing in the most sensitive region at this playback level. Re-dither or re-quantisation can be very audible.
At 20-bits, it's OK, and you can do 5 processes before it touches threshold of hearing.
24-bits: "ludicrously over specified".
He mentioned some room noise data from Louis Fielder that's relevant to this discussion.
He showed a plot of the Shannon space for CD, 52kHz 11-bit (which, with optimal noise shaping, matches human perception perfectly), and 96kHz 24-bit.
Then he worked through an interesting set of statements which explain why CD is nearly good enough:
Firstly, the noise in even high-res recording is always above that which can be achieved on CD.
Secondly, while it's true that 120dB at 24kHz can't be stored on CD (despite being audible to some people), there's no actual content like that, and it would melt tweeters anyway. A "real world" worst case is something like a cymbal crash, and by about 20kHz even that is below the threshold of hearing.
The preference for higher sample rates is not based on ultrasonic content, but on time resolution and aliasing when conversion goes wrong. All anti-alias filters are brick wall, anti-causal (i.e. pre-ring) which is unnatural. You can hear the difference between different sample rates, even when the tweeter doesn't go above 18kHz and/or your hearing stops at 17kHz.
When upsampling CD content, they are using an "apodizing" filter. How did they develop this? Firstly Peter Craven created controlled pre-response anti-alias filters. These have a null near the Nyquist frequency, which will kill any ringing from the original A/D converter.
The ringing is within the time domain that is masked by the human auditory system.
In listening tests, they taught themselves what filter ringing sounded like by starting with a maximum phase filter. This creates a "glassy, bright" sound, and also a kind of phosphorescence or sparkling on top and around the audio.
Then they carried out listening tests on 12 filters. They're not making public all the changes , but the final choice was decided purely by listening tests, and is substantially minimum phase.
Multi-channel audio.
When it came to DVD-A, the recording industry's priorities were copy protection, copy protection, surround sound, and menus.
DVD-A and SACD had a battle, and the iPod won.
Why did it fail? The studios didn't supply content, and CD is already pretty good. Importantly, CDs (which will sound great in great systems) are bought by ordinary people and played in boom boxes - the format is supported by ordinary people. (my interpretation: Any format that isn't, can't survive.)
They have got something into BluRay by stealth - they made the argument that blind people need to be able to operate BluRay player(s), and on this basis the features exist to navigate BluRay without access to a screen - leaving the possibility of using it as an audio-only carrier.
Height is incredibly important - more so than rear channels. He mentioned ADG 2+2+2.
A problem with 5.1 is that 2 good speakers are better than 5 nasty ones. Also, when it comes to laying a room out, height in the front image is more acceptable than placement of rear speakers, but you can't get it from 5.1.
ABX testing
Someone raised the question of the recent blind testing published in the AES that showed 16/44.1 is enough. He replied that, in a familiar system, it's possible to hear things through the system's own faults. However, it can take days to hear a fault.
He argued that Double Blind Testing is used a lot in psychoacoustics, to measure masking etc, but when you move from perception to cognition, it becomes different. ABX testing can disconnect you from things that are continuous. He presented several anti-ABX slides, listing things that DBT proponents believe are inaudible, suggesting that idea that these things sound the same is ridiculous. e.g. 2 violins never sound the same, why would you expect 2 loudspeakers to ever sound the same?
Perceptual memory breaks ABX. He gave the example of listening to a familiar recording on a new high quality audio system: you might hear new details in the recording that you had never appreciated before. He gave an example of a recording where he discovered there were three guitars, rather than the two he had heard before. Having heard this, whenever he listened to the recording afterwards, on almost any system, he could appreciate that there were three guitars.
So in ABX testing, if there's something you can't hear in A, but then you hear it in B, then you'll often be able to hear it when you listen to A again, because your memory fills it in. There was a real difference, but your perceptual memory fills in the gaps to prevent you hearing it more than once.
Hearing is non-linear, and allows you to concentrate on different things, which are both flaws with ABX testing.
(my own note: he didn't state explicitly, but it occurred to me from what was said, that there's a problem when you don't have a break between A and B - because a break tells your brain "this is (or might be) something different", whereas the lack of a break tells you brain "this is more of the same" and may impair the audibility of differences. FWIW I always ABX with breaks between samples, but I'm not sure everyone does - certainly the hardware ABX boxes try to prevent it).
After the official Q&A, I talked to both Bob, and his engineers.
One interesting comment is that the industry wanted HD-DVD or BluRay mainly because China is building DVD players too cheaply, and they need a new format so we can profit from player sales (patents?) again.
The minimum-phase-ish Apodizing filter is used like this:
44.1k > zero pad > 88.2k > apodizing filter @ 22k
In the active loudspeaker cross over filters, they're still using linear phase filters - they sound better. The engineer made the point that even the apodizing filter is still linear phase ish up to ~18kHz.
I asked Bob about his claim that the problem with linear phase brick wall filters is that you hear the wavefront early, and it can damage binaural timing cues. I asked "how can you hear it - it's at 22kHz?!". He said stop thinking about the ear as a linear filter - it's not - it's a wavefront detecting device. I said the cochlear models I'd seen still didn't respond to 22kHz pre-ringing. He said if so, they're missing something - he believed some accurate models would respond. He pointed out that the ear is a very poor performer in many ways, but an excellent difference detector.
I asked how much of the audible difference is due to distortion or non-linearity in the equipment. He said "some, it could be a factor" and then smiled to himself. (my note: I'm not sure how to interpret that smile!). Then he repeated that the ear is a wavefront detector, and ear models which show a simultaneous parallel response should show it.
He also said that he shouldn't say it, but that CD, properly sampled, is pretty much as good as high-res audio. Close enough that it doesn't matter. High Res has slightly better time resolution, but not much.
Talking to his engineers, they said most tests are sighted, or sometimes single-blind. He believes that the A/D anti-alias filter has a bearing on which reconstruction filter sounds best.
Talking to an engineer from another audio company, he said they rarely carry out even single blind tests. They don't believe they're biased because they don't know what to expect to hear - e.g. they change a filter, and it might cause a change in the stereo image, or the change in the sound of a triangle. Different filter = different space perception. While the designer claimed to be able to hear some differences, he said the most subtle changes were only audible to a few golden ears in the company.
That's the lot. If you want it from the horse's mouth, download the mp3 linked in the previous post!
I could add some fairly obvious comments to the ABX section, but I'm sure I don't need to bother.
Cheers,
David.
P.S. some relevant Google results:
Review of CD player with Apodizing filter:
http://www.meridian-audio.cz/test/meridian...hoice_11_08.pdf (http://www.meridian-audio.cz/test/meridian-g08-2-hifi_choice_11_08.pdf)
Peter Craven's AES paper on Apodizing filters:
http://www.aes.org/e-lib/browse.cfm?elib=12992 (http://www.aes.org/e-lib/browse.cfm?elib=12992)
I could add some fairly obvious comments to the ABX section, but I'm sure I don't need to bother.
It appears to me that he has totally missed the point of ABX testing. ABX can't prove that a difference doesn't exists or can't be heard. It merely tests the hypothesis that a particular listener under specific conditions is able to hear a difference with statistically significant probability.
P.S Thanks for passing on this information, and for all of the contributions you have made to HA over the years.
24-bits: "ludicrously over specified".
Thanks for the link. I did listen to the AES UK recording, and recall that he said that 96kHz/24-bit was ludicrously overspecified, and repeated both the sampling rate and bit depth a minute of two later.
That ties in to the 52 kHz sampling rate he later mentioned (implying no more the 26 kHz should be audible to humans at even the loudest level, even though it's unlike anything in music). If that's true, there's 44 kHz of excess sampling frequency (22kHz excess bandwidth) at 96kHz, which is pretty ludicrous. 24-bit is also OTT, though it's often sensible in computers to use an integer number of bytes per sample per channel.
Shannon space seems to encapsulate the 'gamut' of human audition, analogous to the gamut of colour space for human vision.
He mentioned some room noise data from Louis Fielder that's relevant to this discussion.
I hadn't picked up the name, so that's handy if I want to look it up. It sounded like that was from among the quietest listening rooms possible, though the graph he referred to doesn't show up too well on the mp3!
It could well be true but be irrelevant to "real-world" listening rooms with peak level set to about as loud as a human can bear, or ultra-quiet rooms with more modest listening levels. In either of those circumstances, it seems that 16-bit is good enough for the delivery format (though digital reconstruction filters in DACs or digital speakers might as well be implemented in higher resolution to keep any further dither to lower levels.
I thought the beginning was a useful introduction for those not too familiar with speaker design, where he discussed the improvements possible by using active speakers with a separate amplifier dedicated to each driver and the improvements in sensitivity, frequency response, bass extension, loudness, thermal management and protection that could provide compared to traditional passive designs where one has to forego sensitivity, attempt to match all driver impedances and provide power-capable components in the passive crossover.
2008-12-09 AES UK Bob Stuart Meridian
ABX testing
Someone raised the question of the recent blind testing published in the AES that showed 16/44.1 is enough. He replied that, in a familiar system, it's possible to hear things through the system's own faults. However, it can take days to hear a fault.
He argued that Double Blind Testing is used a lot in psychoacoustics, to measure masking etc, but when you move from perception to cognition, it becomes different. ABX testing can disconnect you from things that are continuous. He presented several anti-ABX slides, listing things that DBT proponents believe are inaudible, suggesting that idea that these things sound the same is ridiculous. e.g. 2 violins never sound the same, why would you expect 2 loudspeakers to ever sound the same?
Perceptual memory breaks ABX. He gave the example of listening to a familiar recording on a new high quality audio system: you might hear new details in the recording that you had never appreciated before. He gave an example of a recording where he discovered there were three guitars, rather than the two he had heard before. Having heard this, whenever he listened to the recording afterwards, on almost any system, he could appreciate that there were three guitars.
So in ABX testing, if there's something you can't hear in A, but then you hear it in B, then you'll often be able to hear it when you listen to A again, because your memory fills it in. There was a real difference, but your perceptual memory fills in the gaps to prevent you hearing it more than once.
Hearing is non-linear, and allows you to concentrate on different things, which are both flaws with ABX testing.
(my own note: he didn't state explicitly, but it occurred to me from what was said, that there's a problem when you don't have a break between A and B - because a break tells your brain "this is (or might be) something different", whereas the lack of a break tells you brain "this is more of the same" and may impair the audibility of differences. FWIW I always ABX with breaks between samples, but I'm not sure everyone does - certainly the hardware ABX boxes try to prevent it).
Actually, the idea that "the hardware ABX boxes try to prevent (breaks between samples) is quite false. If you check the Clark JAES article about ABX you will find that the ABX Comparator schematic published there includes a means for adjusting the break between samples over a wide range. Now some other hardware ABX boxes may not have implemented this feature, but that would be the responsibility of their constructors.
I find it interesting that Mr. Stuart presented the above as being a fact, not a personal theory that needs to be confirmed by testing. That pretty well trashes his Scientist's Card right there! :-(
The second external problem with what he said is that he said his theory breaks just ABX testing. If his theory were correct it would break many forms of testing, including ABC/hr and sighted evaluations. But he didn't say that so that pretty well trashes his backup Scientit's Card.
Basically, we're seing evidence of what may well be Stuart's personal vendetta against ABX. Or, it might be a wish or a dream that he tells himself to reconcile the fact that ABX pretty well rains on many of his favorite parades. If ABX is right, then its hard to escape the idea that Stuart has been spinning his wheels, and his customer's wheels as well.
Now, to get inside Stuart's theory about ABX. The first problem is that Stuart's theory presumes that all ABX tests involve quick switching and that there is always a "lack of a break". This is as false as the similar flase theory that ABX tests all involve short snippets of sound. Anbody who has hands-on experience with ABX testing should know better from personal experience. The switching and the size of the audible samples can be whatever the listener wants them to be.
Another false claim is quoted above:
"He presented several anti-ABX slides, listing things that DBT proponents believe are inaudible, suggesting that idea that these things sound the same is ridiculous. e.g. 2 violins never sound the same, why would you expect 2 loudspeakers to ever sound the same."
I don't believe that there is a widespread belief among ABX proponents that any 2 different loudspeakers sound the same. So, we have a clear case of a straw man argument. I find this level of posturing to be completely dismaying coming from a person of Sturat's stature.
"(Stuart) gave the example of listening to a familiar recording on a new high quality audio system: you might hear new details in the recording that you had never appreciated before. He gave an example of a recording where he discovered there were three guitars, rather than the two he had heard before. Having heard this, whenever he listened to the recording afterwards, on almost any system, he could appreciate that there were three guitars."
This is a fact, but its not just about ABX, its about listener training and it affects just about any kind of listening test you want to do. That a person would hear 3 guitars on a poor system where the sound of 3 guitars might be more masked than on a good one is just an example of listener bias. It is a natural consequence of how human perception works. What it really says is that the idea that listeners should always use music that they are familar with may not always be the best idea.
This is a common method of argumentation that highly biased people use. They come up with some problem that may be quite general, bnut they ascribe it only to something that they want people to distrust. It's like having two girls named Sandy and Julie in a beauty contest. They are both a little zaftig, but someone repatedly says that Sandy is fat. Guess which girl looks at least a little more plump to most people?
Then Stuart further puts his head in the same noose by saying the following:
"Hearing is non-linear, and allows you to concentrate on different things, which are both flaws with ABX testing."
Obviously this possibly real problem (the description is so vague I can't tell whether it is fish or fowl) that he is mentioning potentially afflicts a wide range of listening tests. It might affect all of them.
Bottom line is that there are a lot of inherent problems with just about any kind of listening test. Listening test involve humans, and humans are messy creatures as we are all reminded several times a day. So what are we to do, resolve not to ever do them?
It appears that Bob Stuart is not lighting candles against the darkness created by invalid listening tests but running around and trying to blow out some of the more effective candles that are currently burning.
This report makes it look like Bob Stuart is clearly on the warpath against ABX, and ruining his own reputation for objectivity and honesty by spewing trashy rhetoric, it would seem.
I could add some fairly obvious comments to the ABX section, but I'm sure I don't need to bother.
It appears to me that he has totally missed the point of ABX testing. ABX can't prove that a difference doesn't exists or can't be heard. It merely tests the hypothesis that a particular listener under specific conditions is able to hear a difference with statistically significant probability.
P.S Thanks for passing on this information, and for all of the contributions you have made to HA over the years.
I just downloaded the MP3 that people seem to be talking about from http://www.aes.org/sections/uk/meetings/AE...ecture_0812.mp3 (http://www.aes.org/sections/uk/meetings/AESUK_lecture_0812.mp3)
At what MM:SS offset in this file might the comments about ABX be found?
I haven't listened to the mp3 (I was there!), but my report is in chronological order - so it's towards the end-ish.
"ABX" was my heading. I think he started off saying "DBT", but did move on to saying ABX.
The stress on "no gaps" is mine - I noted as such.
IMO some of the excuses raised against ABX are just pure fluff. As you've noticed, some are classic "straw man" argument tactics. Interestingly, the vocal part of the audience didn't seem to need any pushing to accept his opinion. There were plenty of grunts of agreement.
FWIW his own engineers weren't so ardently anti-DBT, they just felt that generally they didn't have time to do them (or it wasn't the best use of their time), and that any listener bias would be removed/averaged by the amount of listening, and number of people listening, that goes on. This would work OK if the listening was truly independent and blind, but as I noted, they "occasionally" do single blind testing. I assume everything else is sighted.
Cheers,
David.
I just downloaded the MP3 that people seem to be talking about from http://www.aes.org/sections/uk/meetings/AE...ecture_0812.mp3 (http://www.aes.org/sections/uk/meetings/AESUK_lecture_0812.mp3)
At what MM:SS offset in this file might the comments about ABX be found?
85:13 in MM:SS (or 1:25:13 in H:MM:SS). The whole Q&A session started at about 1:24:00. Before this, Bob had only mentioned in passing the recent paper about the audibility or otherwise of a 44100Hz/16-bit stage inserted in a high-end playback system.
I just downloaded the MP3 that people seem to be talking about from http://www.aes.org/sections/uk/meetings/AE...ecture_0812.mp3 (http://www.aes.org/sections/uk/meetings/AESUK_lecture_0812.mp3)
At what MM:SS offset in this file might the comments about ABX be found?
85:13 in MM:SS (or 1:25:13 in H:MM:SS). The whole Q&A session started at about 1:24:00. Before this, Bob had only mentioned in passing the recent paper about the audibility or otherwise of a 44100Hz/16-bit stage inserted in a high-end playback system.
Thanks for the quick reply. I found what I was looking for with your help. The active speaker parts of the lecture seemed to be more helpful.
FWIW his own engineers weren't so ardently anti-DBT, they just felt that generally they didn't have time to do them (or it wasn't the best use of their time),
I suspect that were Stuart's own engineers to do proper DBTs, any number of projects related to digital coding and decoding would be scrapped. That could unfavorably impact their bottom line.
and that any listener bias would be removed/averaged by the amount of listening, and number of people listening, that goes on. This would work OK if the listening was truly independent and blind, but as I noted, they "occasionally" do single blind testing. I assume everything else is sighted.
Single blind = an intentionally defective DBT, whose failings have been known since the early 1800s and "Clever Hans".
I note that Stuart repeated the long term listening test straw man argument, did in fact as previously said entertain the audioence with this AFAIK newly-minted perception versus cognition straw man, the ABXer's expect speakers to sound the same straw man, the ABX tests fail for any test involving memory straw man, the listener training only affects ABX tests straw man, the straw man about the nonlinearity of the human ear only affecting ABX tests, the there are no such things as long term DBTs straw man, the straw man arugment that ABX is far more flawed than sighted evaluations, the focussed attention affects only DBTs straw man, and the straw man argument that ABX tests uniquely presume that people listen to the same music the same way every time.
Stuart summarizes all of the above by saying that "it (ABX) is of dubious interest". Pretty strong stuff. :-(
Not bad, for 8 minutes of work! ;-)
85:13 in MM:SS (or 1:25:13 in H:MM:SS). The whole Q&A session started at about 1:24:00. Before this, Bob had only mentioned in passing the recent paper about the audibility or otherwise of a 44100Hz/16-bit stage inserted in a high-end playback system.
...which was sparked by his 2004 JAES paper on Coding for High Resolution Audio Systems.
So there is definitely some history here.
I'm glad Dr. Stuart admits that CD can be practically 'good enough'. I'm glad he has also (like in that 2004 paper) asserted contra the audiophile reactionaries that CD CAN resolve below the LSB, and that it CAN resolve events in time finer than the sample rate. But shame on him for that loudspeaker difference strawman.
I'm glad Dr. Stuart admits that CD can be practically 'good enough'. I'm glad he has also (like in that 2004 paper) asserted contra the audiophile reactionaries that CD CAN resolve below the LSB, and that it CAN resolve events in time finer than the sample rate. But shame on him for that loudspeaker difference strawman.
I was thinking about the potential profit that Meridian is foregoing by wasting so much effort on digital player tweaks that won't show a difference in the ABX tests Dr. Stuart obviously despises.
I was thinking about the potential profit that Meridian is foregoing by wasting so much effort on digital player tweaks that won't show a difference in the ABX tests Dr. Stuart obviously despises.
Are you joking? I assumed that all such journeys away from the scientific method (especially by men of science) are profit driven.
FWIW I think the real tragedy is that all the things the tiny audiophile minority has become obsessed with have prevented us from bringing real, comparatively easy, genuine improvements to reproduced audio. One that Bob mentions in the lecture is height information in recordings. I've heard ambisonic recordings (of real live music, not aeroplanes!) decoded with height information (i.e. with two extra, high, channels added at the front) and the change in the perceived sound stage is stunning - both in height and depth.
Maybe if the industry was chasing after real improvements like that - i.e. ones that anyone can hear in any kind of test(!), they'd be doing better?
Of course, that kind of improvement needs the whole industry to join the party and deliver suitable recordings. No sign of that.
So where can money be made? Why, by offering an improvement in the reproduction of the 2-channel 44.1kHz 16-bit audio that's available by the cart load. Maybe an (inaudible?) improvement in this can be more successfully marketed than a genuine improvement which is incompatible with all existing recordings.
Cheers,
David.
I was thinking about the potential profit that Meridian is foregoing by wasting so much effort on digital player tweaks that won't show a difference in the ABX tests Dr. Stuart obviously despises.
Are you joking?
no
I assumed that all such journeys away from the scientific method (especially by men of science) are profit driven.
They might be, but actually represent a failure of leadership vision.
FWIW I think the real tragedy is that all the things the tiny audiophile minority has become obsessed with have prevented us from bringing real, comparatively easy, genuine improvements to reproduced audio.
Don't blame the audiophiles, they are sheep, the victims. They spend money and get nothing for it. IMO, they are educated into this dysfunctional behavior by the Atkinsons and Stuarts of the world.
One that Bob mentions in the lecture is height information in recordings. I've heard ambisonic recordings (of real live music, not aeroplanes!) decoded with height information (i.e. with two extra, high, channels added at the front) and the change in the perceived sound stage is stunning - both in height and depth.
Yes, but what amibsonics does is demonstrable as at least making an audible difference in an ABX test.
Maybe if the industry was chasing after real improvements like that - i.e. ones that anyone can hear in any kind of test(!), they'd be doing better?
Exactly.
Of course, that kind of improvement needs the whole industry to join the party and deliver suitable recordings. No sign of that.
I'm not sure that Ambisonics is sufficiently consumer-oriented to make it in the mainstream.
So where can money be made? Why, by offering an improvement in the reproduction of the 2-channel 44.1kHz 16-bit audio that's available by the cart load. Maybe an (inaudible?) improvement in this can be more successfully marketed than a genuine improvement which is incompatible with all existing recordings.
Sturat seems to have some good ideas about active loudspeakers. I'm not sure how innovative they are at this time.
Maybe if the industry was chasing after real improvements like that - i.e. ones that anyone can hear in any kind of test(!), they'd be doing better? Of course, that kind of improvement needs the whole industry to join the party and deliver suitable recordings. No sign of that. So where can money be made? Why, by offering an improvement in the reproduction of the 2-channel 44.1kHz 16-bit audio that's available by the cart load. Maybe an (inaudible?) improvement in this can be more successfully marketed than a genuine improvement which is incompatible with all existing recordings
Enough people have been burned (or at least nonplussed) at the various multichannel schemes over the last 40 (!) years, that they simply do not trust people when they say that it really is the way forward. I mean, look at how many audiophiles claim perfect, "3D" imaging from two speakers... that said, there is occasionally some controversy that erupts when some magazine or another devotes more attention to multichannel instead of 2 channel, so I dunno.
Certainly, much of the effort exerted by high-end audio manufacturers is based on de-commoditizing what would otherwise be an inexpensive commodity product. Blind testing threatens to explicitly point out that their products are essentially overpriced commodities. And I suppose many people refuse to believe that their audio equipment (and their music!) are, more or less, commodities - they want to be a part of something special. I think
that's another underlying emotion keeping this whole machine going.
If y'all don't mind, I'd like to fork off the multichannel discussion into a separate thread. (http://www.hydrogenaudio.org/forums/index.php?showtopic=68834)
Enough people have been burned (or at least nonplussed) at the various multichannel schemes over the last 40 (!) years, that they simply do not trust people when they say that it really is the way forward. I mean, look at how many audiophiles claim perfect, "3D" imaging from two speakers...
How many of those people heard properly set up multichannel systems, playing recordings that really were intended to sound like music being playing in a real space, in real time?
That said, there is occasionally some controversy that erupts when some magazine or another devotes more attention to multichannel instead of 2 channel, so I dunno.
Taking the luddite proclivites of the 'high end' magazine readers as indicative of audio preferences, is not a safe bet. The 'mass market' magazines have pretty much embraced multichannel setups. It's where the money is too.
This report makes it look like Bob Stuart is clearly on the warpath against ABX, and ruining his own reputation for objectivity and honesty by spewing trashy rhetoric, it would seem.
The reference to different violins sounding different (an idea most people would readily accept) supported the assertion to the effect that loudpeaker systems even though the same model constructed by the same manufacturer will sound slightly different (something many people might find surprising). As a variation on this idea, if you attempt to A B compare one loudspeaker system model against a different model, that will give you a test result for the particular units involved, but may not be representative of the difference in quality between the models if more examples (actual manufactured units) were compared of each model. So the question can be posed: 'What have you achieved?'. Stuart did not actually explain his reference to loudspeakers not sounding the same; nor did he explicitly pose the question I have just posed. He was after all in the final stretch of his session (answering questions). My interpretation was that he was harking back to a theme he put forward in the main part of his talk. There he made the point that you should approach speaker design from the point of view of fundamental strategies (e.g. using separately amped speakers rather than passive crossovers) and would leave comparative subjective testing till much further down the track. His theme in the talk seemed to be: get the engineering done and take your measurements and do your homework. Use listening tests right at the end after you have a solid foundation. The listening test is the ultimate finessing of the system, e.g. to choose between different filtering that on paper should be equallly effective but which for reasons not yet understood by science do sound different.
No doubt
some listening tests would be done along the way, to establish/confirm the new engineering was heading in a direction likely to give tangible benefits.
My take when listening to the mp3 of the talk was that Stuart recognised the value of ABX in some contexts but not others. Some actual words (if my transcription is accurate):
[blockquote]You know those kind of tests are used a lot in psychoacoustics when you're trying to measure the capability of the hearing system in response to certain kinds of stimuli. ... As soon as you get from perception to cognition, it becomes different. ... Maybe it's after several minutes that you either hear defects or things in the recording. So ABX testing I think is very very good for what is for but it is seriously flawed for at least that reason and there are other reasons.[/blockquote]The talk struck me on the whole as open and balanced. I don't think it was vehemently anti ABX, but it did point out some characteristics of human cognition that could interfere with the brain's ability to perform consistenty and 'accuarately' in an ABX setting, for some kinds of deficiencies.
It is an amazing thing that once having heard something subtle in a musical recording with the help of a very good speaker system, one can thereafter hear it even on an inferior system. But had one only ever been listening with the inferior system, the subtle thing might always have remained unnoticed. I would imagine though that despite this, the inferior system would probably still sound inferior, or at least 'different', for a
careful listener, in an ABX test. But then what is the next logical step for a loudspeaker engineer in improving the design of the inferior speaker system? How has the ABX test really helped the engineer?
in case anyone ever visits here again, the full path to the mp3 of Bob Stuart's 2009 talk is now
http://www.aes-media.org/sections/uk/meeti...ecture_0812.mp3 (http://www.aes-media.org/sections/uk/meetings/AESUK_lecture_0812.mp3)