Sections 2, 6, 17 of 15 Haydn_ String Quartet In D, Op_ 76, No_ 5 - Finale - Presto + cues for AES Conference Paper 9174: Audibility Of Typical Digital Filters
"
We analysed the percent-correct scores for
each of sections 1-17 across all conditions, and found
that some sections yielded a higher ratio of correct
results than others. For example, sections 2, 6 and
17 gave correct-to-total ratios of 0.714, 0.710 and
0.769 respectively
"
[attachment=8064:15_Haydn...ection_2.flac]
[attachment=8065:15_Haydn...ection_6.flac]
[attachment=8067:15_Haydn...ction_17.flac]
Left channel of third sample:
(http://i60.tinypic.com/ilvt4l.png)
Spectral images a-plenty, possibly from low quality resampling.
Doesn't this look like dsd noise and maybe a hefty imaging problem resulting from a faulty dsd -> PCM conversation?
D'oh!
So much for getting access to all the bits.
Arny, what is the source of these samples ? Is it sure that they are identical to the original versions on the 2l.no website ?
ps: I've just checked with the original I've had for months, and your version is about 2 dB softer !
Doesn't this look like dsd noise and maybe a hefty imaging problem resulting from a faulty dsd -> PCM conversation?
... or a faulty PCM > DSD conversion.
Arny, what is the source of these samples ? Is it sure that they are identical to the original versions on the 2l.no website ?
ps: I've just checked with the original I've had for months, and your version is about 2 dB softer !
I compared the last sample (cue 17) with a high resolution (96k stereo) version I downloaded several years ago, and yes the cue 17 version is about 1.7dB softer.
[On another matter, I seem to be getting errors on some flac downloads. The error manifested itself in the decoded waveform for 15_Haydn__String_Quartet_In_D__Op__76__No__5_-_Finale_-_Presto___cues_section_17.flac by way of silence for a short segment of about 21mS commencing 1.834mS into the file. Probably an error at my end. I am querying this at post #211 (http://www.hydrogenaud.io/forums/index.php?s=&showtopic=101082&view=findpost&p=881350) of the
FLAC 1.3.0 has been released thread.]
Doesn't this look like dsd noise and maybe a hefty imaging problem resulting from a faulty dsd -> PCM conversation?
... or a faulty PCM > DSD conversion.
The L2 record company uses Weiss Saracon for SRC, which is a very well known, highly regarded and reliable professional tool. This does not guarantee that something has gone wrong though. I've noticed before that L2 uses a default setting in which the output is 0.2 dB lower in order to prevent overloads. This does make sense, but it should be noted when carefully comparing the original DXD against the SRC'd versions.
Welcome to the complimentary download area (http://www.2l.no/weiss/) brought to you by 2L and Weiss Engineering.
We invite you to download high resolution files for free and play them on your Weiss equipment. The music has been recorded in DXD (352.8 kHz sampling rate) and subsequently converted to 192kHz / 24Bit using the Weiss Saracon Sampling Rate Converter software.
[/size]
Arny, what is the source of these samples ? Is it sure that they are identical to the original versions on the 2l.no website ?
ps: I've just checked with the original I've had for months, and your version is about 2 dB softer !
I compared the last sample (cue 17) with a high resolution (96k stereo) version I downloaded several years ago, and yes the cue 17 version is about 1.7dB softer.
From the paper:
"
3Please note that 2L also offers a free sampler version of this which is not identical to the one we used; the le \2L53SACD 04 stereo 192kHz.
ac" available at the URL http://www.2l.no/hires/ (http://www.2l.no/hires/) has the level raised approximately 2 dB so as to peak at -1 dB.
"
This one is the pay-for version.
Rather than add a new thread, I'll use this one for other uploads related to the discussion of the AES conference paper: "The audibility of typical digital audio fllters in
a high-delity playback system"
[attachment=8070:Meridian...eaker_FR.png] - FR & implulse response of Meridian DSP7200SE active speaker system
We invite you to download high resolution files for free and play them on your Weiss equipment. The music has been recorded in DXD (352.8 kHz sampling rate) and subsequently converted to 192kHz / 24Bit using the Weiss Saracon Sampling Rate Converter software.
DXD rate recorded is not very common imho. If some ADC gives out this rate it most lkely converts it internally from its multibit bitstream. Also possible they record in dsd, open it with a program that converts it to DXD like Pyramix but see it recorded in DXD.
Still strange to see that pattern.
DXD rate recorded is not very common imho. If some ADC gives out this rate it most lkely converts it internally from its multibit bitstream.
That implies that DXD is singlebit, but it is in fact multibit as well.
DXD was created because DSD is pretty much impossible to work with, it is 24-bit @ 352.8kHz. DSD runs at an even higher rate, 2.8MHz, but apparently there are people working with DSD128, DSD256 and DSD512, the latter runs 1-bit at a whopping 22.6MHz.
Because, obviously, bigger numbers are better.
Spectral images a-plenty, possibly from low quality resampling.
I've been looking at this. I'm not sure this is the explanation.
For one thing, the few other samples from 2L I tried...
http://www.2l.no/hires/index.html (http://www.2l.no/hires/index.html)
...don't have this problem. The Britten is sufficiently similar that you'd expect it to exhibit the same problem, but it doesn't at all. I'm making the assumption that all the files went through the same resampling, and nothing went uniquely wrong with the Haydn. I'm not sure those assumptions are valid.
For another thing, the DXD 24/352.8 version has just the same "spectral images". They claim the DXD version is the master, which the other versions were created from. Well, it's there on the "master", if that's what it is.
Most convincingly, look carefully at the spectrum, and listen carefully to the recording (replaying the 192kHz version at 48kHz and then 96kHz, so you can hear those "images"). It sounds rather metallic (just like lousy resampling, actually!) but it rings along
after the loud sections. It doesn't start immediately when there's a loud note (there's a short delay), and it rings on afterwards (it dies out with the reverb). I think it's something in the recording venue rattling along with the music. I bet it's some little metallic thing that's inaudible in real life, that becomes audible when you slow the playback down (just like bats). I suspect it's just about audible at normal playback speeds to some people, but without a "clean" version to compare with, it's hard to be sure.
Does anyone know full details about DXD recording? I've read the 2l and lindberg web pages, but I don't feel I know exactly what's happening. It's strange the DXD version has -45dBFS ultrasonic noise, but that's nothing compared to the DSD version with (in the same bandwidth) -26dBFS ultrasonic noise. If DXD is edited DSD, how is it so much better? If DXD is the native recording format, does the ultrasonic noise come from the converter's noise shaping, or elsewhere?
Cheers,
David.
P.S. Some of the 96kHz downconversions are slightly strange. The top ~3kHz are empty. It's not even a gentle roll off - it goes off a cliff at 45kHz. Not the Haydn though.
P.P.S. As already mentioned we're still not getting "the original bits" - because there's no DXD 5.1 download.
2L says says the Haydn tracks were 'produced in DXD' but they also offer DSD versions -- is going from DXD to DSD really something that is done? I would expect to see the production chain go the *other* way (DSD native recording -->DXD production)
is going from DXD to DSD really something that is done?
Pretty often yes. People still want DSD (lots of marketing pressure), and only reason to do DXD is that either source or main target is DSD.
From memory, DXD is the editing format for DSD. Seems so per wikipedia (http://en.wikipedia.org/wiki/Digital_eXtreme_Definition).
cheers,
AJ
Can somene else confirm the relative noise levels? And/or "typical" native DSD noise levels in that bandwidth? It might help to figure out exactly what they've done.
Cheers,
David.
Here are typical DSD noise levels. Not sure how much they modulate with the signal, but shouldn't be a lot:
(http://prosound.ixbt.com/education/dithering-noise-shaping/images/dr.png)
Here are typical DSD noise levels. Not sure how much they modulate with the signal, but shouldn't be a lot:
(http://prosound.ixbt.com/education/dithering-noise-shaping/images/dr.png)
Point of order: I think those are unprocessed FFTs with constant bandwidth in terms of frequency (and that isn't said, either) so I don't know what they mean until they are reformatted into fractional octaves - 1/3, 1/6 or 1/12.
Arnie, these graphs are for David. Sure, they are regular FFTs which are always linear in frequency. I don't have equivalent graphs with octave-integrated energy.
Do you know the FFT size/length (and actual sample rate at which it was calculated) and window type?
Thanks,
David.
No, but this can be inferred from the graph: SR = 96 kHz, FFT size = 16384 (I could be off by 2 times).
If the "PCM 16 44" is with triangular dither then the linear spectral density [Vp/sqrt(Hz)] should be about 10 dB lower than the amplitude shown in the graph.
From memory, DXD is the editing format for DSD. Seems so per wikipedia (http://en.wikipedia.org/wiki/Digital_eXtreme_Definition).
Yes, exactly -- one of several flavors of PCM devised so that you can actually do digital production on DSD sources while staying at ultrahigh sample rates.
But if in fact people are now recording *directly* in DXD (which, again, is a flavor of PCM), then converting to *DSD* for DSD/SACD fans -- well, I would be amused but not terribly surprised at such patent nonsense.
If the "PCM 16 44" is with triangular dither then the linear spectral density [Vp/sqrt(Hz)] should be about 10 dB lower than the amplitude shown in the graph.
I don't think so, I can repro this graph in iZotope RX or other spectrum analyzers. Perhaps you forgot about variance, the attached spectrum is clearly time-averaged.
I don't think so, I can repro this graph in iZotope RX or other spectrum analyzers. Perhaps you forgot about variance, the attached spectrum is clearly time-averaged.
These spectrum analyzers don't display density, they assume distinct peaks (tones) which noise is not and plot amplitudes instead. Noise may be over- or underestimated depending on all the parameters (window, fft size, Fs ...).
Noise density of 16/44.1 with TPDF should be around -137 dB. (I was too tired yesterday, so I was a few dB off..)
For example in the 1/3rd octave 4 kHz band this noise density would add up to -107 dB.
For DSD 2.8 it would be
roughly -127 dB. Of course this gets worse. Absolutely worst would be the 30 kHz band with the tons of noise DSD 2.8 is producing up there.
Ah, gotcha. Sure, I actually don't know any spectrum analyzers that display density. Most of them display amplitude of FFT bins, nothing more.
I wonder if the strange notches in the DSD noises have special meanings? They don't look like traditional equal loudness contours and other noise shaping curves as shown in Alexey's research.
http://audio.rightmark.org/lukin/dither/dither.htm (http://audio.rightmark.org/lukin/dither/dither.htm)
I wonder if the strange notches in the DSD noises have special meanings? They don't look like traditional equal loudness contours and other noise shaping curves as shown in Alexey's research.
http://audio.rightmark.org/lukin/dither/dither.htm (http://audio.rightmark.org/lukin/dither/dither.htm)
I bet it's the response of the noise shaping filters. Most digital filters have massive notches in their stop-bands where the "zeros" are. They're assumed to be irrelevant, because though the response theoretically drops away to -infinity at these points, it's always falling from an already inaudible -100dB or so.
To put it another way: when digital filters are designed to attenuate by, say, 100dB, that's the minimum attenuation; -100dB is the
most signal that will get through. At the "zeros" (i.e. the frequencies at which the stop band response is pinned down), it will be "almost" infinitely attenuated.
Graphs don't normally have axes that let you see this, but it's usually there.
Cheers,
David.
Does anyone here have experience with string quartet sound pressure levels ?
The Stuart paper states:
For a system gain of 75dB, the loudest peak passage was measured as 102 dB SPL at the listening position, somewhat lower than the level we would expect from a live performance at a distance of 3 m. This level was chosen for comfort, and because it was high enough for details to be audible but also low enough that 16-bit RPDF dither would be inaudible at the listening position
I have read another study about "Loudspeaker Simulation of a String Quartet for in situ Listening Evaluation"
http://www.acoustics.asn.au/conference_pro.../Papers/P2a.pdf (http://www.acoustics.asn.au/conference_proceedings/ICA2010/cdrom-ISRA2010/Papers/P2a.pdf)
At the end of the recording session, the A-weighted, peak sound pressure level was measured with a B&K 2260 sound level meter, fast setting, approximately 4m in front of the quartet. This measurement (78dB) was used to roughly calibrate the playback level during listening evaluation.
While I can accept differences between string quartets, measurements and acoustics, the difference between 102 dB and 78dB (both peak) is very large. It indicates to me that the Stuart test might have played the recording at a higher than realistic level.
(http://prosound.ixbt.com/education/dithering-noise-shaping/images/dr.png)
(http://dsd-guide.com/sites/default/files/images/clip_image004.gif)
(supposedly from Andreas Koch's DSD seminar powerpoint presentation)
edit: previous image added on top for better comparison
edit: plotting SNR as a flat line vs. the noise spectrum of DSD noise, (dis)ingen(u/i)ous!
Compared with my Sony PCM-D100 test I am pretty sure the graph Alexey provided is the more accurate one.
http://www.hydrogenaud.io/forums/index.php...ost&id=8068 (http://www.hydrogenaud.io/forums/index.php?act=attach&type=post&id=8068)
Does anyone here have experience with string quartet sound pressure levels ?
Report of 95 dB 'sound power' for a string quartet in an 'auditorium' here:
http://www.zainea.com/hall.htm (http://www.zainea.com/hall.htm)
Does anyone here have experience with string quartet sound pressure levels ?
Lots...just with my ears.
The only time I've ever been 3m away, would be at recitals in the practice rooms outside the main concert hall. 3m from the front row to performers on the main stage...no way!
Next time I'm in such a scenario I'll do some A weighted measurements with my Sony Z1 using Audiotools and an external Dayton mic.
102db seems incredibly loud for such a scenario. Now I'll have to figure out a way to get them to play some Haydn.
cheers,
AJ
The DSD graphs that I posted are just one typical example. Noise floor of DSD depends on the encoder used. The noise shaping algorithm of some encoders creates notches in the noise spectrum, others are more flat. Here's a comparison of few encoders: A Comparison of DSD Encoders & Decoders (http://archimago.blogspot.ru/2014/04/analysis-comparison-of-dsd-encoders.html).
It should also be noted that because DSD is not a properly dithered quantization, the noise floor varies with the signal and contains some amount of nonlinear distortion (but is still below the floor of 16-bit TPDF-dithered quantization).
The DSD graphs that I posted are just one typical example. Noise floor of DSD depends on the encoder used. The noise shaping algorithm of some encoders creates notches in the noise spectrum, others are more flat. Here's a comparison of few encoders: A Comparison of DSD Encoders & Decoders (http://archimago.blogspot.ru/2014/04/analysis-comparison-of-dsd-encoders.html).
(http://dsd-guide.com/sites/default/files/images/clip_image004.gif)
Thanks. After seeing your link it is very clear that this graph is misleading because it implies noise floor of all 24-bit PCM formats below 20kHz are higher than DSD64.
[attachment=8083:DAD_AX24.png]
http://www.digitalaudio.dk/AX24-ADDA-Converter.1492.aspx (http://www.digitalaudio.dk/AX24-ADDA-Converter.1492.aspx)
Edit: http://www.computeraudiophile.com/f13-musi...rces-dxd-20021/ (http://www.computeraudiophile.com/f13-music-downloads-and-streaming/sources-dxd-20021/) (reg. AX24)
Thanks but I am referring to the theoretical spec, not actual AD/DA performance. Just like RMAA can generate 24-bit signal with -145dBFS noise and -180dB noise shown in spectrum graph but no current AD/DA can achieve this performance.
But now at least I know that the notches in DSD noise is not a "trademark" of DSD because it is only Korg's characteristics.
The DSD graphs that I posted are just one typical example. Noise floor of DSD depends on the encoder used. The noise shaping algorithm of some encoders creates notches in the noise spectrum, others are more flat. Here's a comparison of few encoders: A Comparison of DSD Encoders & Decoders (http://archimago.blogspot.ru/2014/04/analysis-comparison-of-dsd-encoders.html).
Hmm... maybe out of subject but, how it would be with 16/44.1 sources (meaning if one would like to ABX CD vs DSD ... how's that done)?