DAC - Deadly Audio Confusion
Jan 12, 2013 at 4:19 PM Post #61 of 70
I think science is somewhat to blame for audiophoolery too. Numbers and specifications get thrown about without relating them to the most important context of all- The thresholds of human perception. If high bitrate lowers the noise floor and higher sampling rates extend the high frequency response, it doesn't matter if humans can't hear it.

There should be an audio equivalent of the USDA daily recommended allowance on the sides of cereal boxes... In parenthesis after the spec should be the threshold point, so folks can instantly see if they're over the line into overkill or not.

In the other subforums having poor measured performance is often a mark of quality. I see several members wetting themselves over an $1100 DAC that has an SNR of 96db (bang on the money for 16 bit CD) but which claims 24 bit resolution. In this day and age I would suggest this is a mark of mediocre engineering when my 1998 Entech can do better. Of course anything that can manage on-spec performance for CD is probably perfectly fine but why pay extra for something that is technologically average at best when my anonymous $90 DAC/headphone amp from ebay can do better and still give the builders a profit. But that is not the worst of it; in Stereophile technically truly awful kit often at several 10s of thousands of $ is treated with bewildering respect often to the puzzlement of the measurer-in-chief john Atkinson. Of course our old friends cognitive biases are hard at play here; put it in a shiny box and you can sell a cold turd for a fortune to some folks...
Back in the past I read a paper (peer reviewed) no idea where that suggested that in normal listening we really cannot tell the difference between 16 and 14 bits. I remember how stunned I was hearing my first CD player a modest Marantz CD63 (1984) and was itself only a 14 bit (4x OS) machine and the clarity and utter lack of perceived noise yet it could only manage about 90db SNR but of course I was a vinyl user back then and 90db was science fiction to LP spinners...
Regardless it seems utterly certain that non-blind tests are of almost zero value. So often (elsewhere) you have two chaps arguing about which CD player has the more extended top end when both have identical FRs. I've even seen folks with a straight face say that the one with the 3db roll off had a more extended top end even after seeing the graphs.... that is some heavy duty self-deception.
What needs to be done is to take a nice clean digital source and gradually degrade it (blindly and randomly) until the difference can be reliably detected. For instance in a JAES paper back in 1979 JVC researchers added low pass filters to music with very high frequency content and concluded that filters at 20k and 18K could not reliably be detected but 16K and 14K filters could by most (not all) listeners, its a test anyone can try. I'm sure i'd fail at anything above 13K now 

Jan 12, 2013 at 6:35 PM Post #62 of 70
I believe that really, any differences between any decent DACs can be chalked up to cognitive bias and improper volume matching.
I have an ODAC and there is absolutely no way ever I would upgrade from it, unless the form factor (something irrelevant to sound) could be improved. Even that is overkill.
Jan 12, 2013 at 7:33 PM Post #63 of 70
Originally Posted by nick_charles /img/forum/go_quote.gif
Back in the past I read a paper (peer reviewed) no idea where that suggested that in normal listening we really cannot tell the difference between 16 and 14 bits.

At a resolution of 14 bits, there is still about 84 dB A-weighted dynamic range (with no noise shaping), and that can indeed be enough at not too high volume levels, and/or listening to dynamically compressed music that never gets really quiet enough not to mask the noise floor. In practice, white noise at 10-15 dB A-weighted SPL seems to be inaudible in a typical not quite silent listening environment (where ambient noise can have an SPL of 30 dB), so it is possible to reach at least 94 dB peaks (with dynamic compression, that is actually loud) before the noise becomes audible even in silent parts of the music.
Jan 12, 2013 at 9:56 PM Post #64 of 70
AES myths workshop video had a short bit on bit depth, where (IIRC) the bit depth goes from 16 bits downwards:
Audio file is here:
Not really a sensitive, formal test, but interesting to hear.
Jan 13, 2013 at 12:19 PM Post #66 of 70
LossyWAV at the standard quality setting averages at about 11 bits, and it sounds just fine (transparent).

Interesting how is that figure calculated ?
Interesting also as that is somewhat below what we normally attribute to a good LP equivalent ( 12 to 13 bits based on noise performance) for me this is a paradox being that it was noise that turned me on to digital back in 84' my (decent if far from  top end) Rega Planar 3/RB300 was sooooooo noisy (and I was not even a headphone listener back then) but if 11 bits is not distracting on a noise basis it suggests my Rega was possibly far worse than 12 bits equivalent.
I remember reading 1970s turntable reviews from the magazine Juilan Hirsch wrote for where he opined that TTs with a noise figure of  better than about 54db were actually good
I found some old measurements figures for high end 70s/80s Turntables - wow and flutter ran between 0.065% and 0.03% and , weighted rumble figures ran from -65db to -85db the best (measured) consistently being the Linn Sondek LP12. -85db is not bad but only the Linn managed that nothing else was better than -78db (technics Sl1000). I suspect my Rega was a good deal worse , it certainly had poor speed stability...
Jan 13, 2013 at 12:35 PM Post #68 of 70
Interesting how is that figure calculated ?

It probably cannot be directly compared to simple linear PCM formats, as the noise level very likely depends on the input signal, i.e. there is more noise in louder/more complex parts, but masking prevents it from becoming audible. Edit: LossyWAV seems to use noise shaping as well.
Jan 15, 2013 at 10:09 AM Post #69 of 70
For those who want to try what lower resolutions sound like compared to 24-bit, here is c.flac from xnor's listening test quantized to various number of bits per sample. It is also downsampled to 44.1 kHz, to reduce the file sizes, and because the quantization noise is more audible at a lower sample rate (more of the same total amount of noise is in the audio band). Since most high sample rate files are also high resolution, it is more realistic to test the low resolutions at 44.1 kHz. The dither used is +/-1 LSB triangular PDF, but with some of the noise power shifted into the top octave (11025-22050 Hz); this was achieved by differentiating 0 to 1.0 LSB uniform distribution noise, adding it to the high resolution signal, and then rounding the result to the nearest integer.
To compare different dither methods, I have the 10-bit file available with standard white noise TPDF dither, uniform distribution dither, and no dither as well:
c_10w.flac - +/-1 LSB triangular PDF dither with white noise
c_10u.flac - +/-0.5 LSB uniform distribution dither with white noise
c_10n.flac - no dithering, only round to nearest integer
All the files can also be downloaded in a single package: dither_test.zip
Jan 15, 2013 at 10:46 AM Post #70 of 70
By the way, the files above show that dithering is not necessarily that important in practice with music, except in very quiet parts, since a complex "real" sound often has enough noise floor to be essentially "self-dithering". The addition of dither could very well just make the sound noisier. Try subtracting the 24-bit file from all the 10-bit versions, and listen to the residual at high volume. Only for the undithered version does it not sound like pure hiss, but that is also the quietest one. Here is the average spectrum of the residual for the standard TPDF dither (green), colored TPDF (blue), uniform dither (yellow), and no dither (red):


Users who are viewing this thread