Skip to main content

Notice

Please note that most of the software linked on this forum is likely to be safe to use. If you are unsure, feel free to ask in the relevant topics, or send a private message to an administrator or moderator. To help curb the problems of false positives, or in the event that you do find actual malware, you can contribute through the article linked here.
Topic: Short samples: 24 bit vs rendered to 16 bits dithered (Read 15927 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

Short samples: 24 bit vs rendered to 16 bits dithered

Source audio: Upright piano, amateur performance, Handel, Gigue No. 1,  first two bars, 9 sec aprox, recorded with a matched pair of Rode NT1-A condensor microphones, connected to a Behring MX 802A mixer, driving a Creative Audigy 4 hub for the analogue to digital conversion, connected to a PC running Windows XP and Asio4All sound drivers, with n-Track studio 4 as the recording software.

The noise of the Behring mixer was a bit high for 24-bit recording using the normal main outputs.  Instead,  I took the audio from the Auxiliary 1 and Auxliary 2 mixes.

The top lid of the upright piano was open.  The microphones were located a few centimetres above the top of the piano, and displaced horizontally about a quarter of the length of the piano on either side of the middle of the keyboard.  They pointed downwards, with a slight tilt towards the centre of the keyboard.

With the Behring mixer turned off, the recording software stereo VU meters registered around -95dB per channel.  With the mixer turned on, but phantom power for the mics off, the VU meters showed around -85dB per channel.  With phantom power on, the VU meters showed around -75dB per channel but this varied with extraneous noises near the recording room, such as heavy traffic.  (The mixer was set to roll off frequences under 75Hz.)

Here is an unmodified extract from the 24-bit stereo 48KHz recording [now in flac format]: [attachment=4521:attachment]
Here is the same extract after bit reduction to 16 bits using triangular dither: [attachment (4313) now deleted to free up my upload quota - this file was not optimally dithered and was unnecessarily noisy]
The format of the modified extract is 24 bits even though the last 8 bits should contain no varying data.

The peak recorded level is -9.5dB.  It was played forte but not fortissimo. Another piece recorded in the same session peaked at -4dB.  [As a comparison, a lossless format audio file on an HD-DVD disk Superman Returns on my Home Theatre PC came in at -3.3dB at the start of chapter 4 (storm at sea) and -9.7dB in chapter 20 (Superman and Lois go flying from the top of the Daily Planet).  This was for Front Left and Front Right.  The chapter 20 music of the movie is very dramatic, and worth listening to in its own right.  It is not background music. ]


The uploads above are made in response to a request made in November 2006 by AndyH in the Listening Tests thread  16 bit vs 24 bit, any samples that work?:

[blockquote]
There are real differences between 16 bit and 24 bit files, just as there are real differences between uncompressed and mp3. As with the differences between wav and mp3, 16 bit vs 24 bit is often difficult to identify by listening.

Test signals are one thing, but has anyone found any 24 bit music recording that can be successfully identified via ABX testing against a properly resample to 16 bit version of same?
[/blockquote]

The bit reduction and triangular dither were carried out using the freeware audio editor, Audacity.

It is a trivial exercise to ABX compare the two versions if a high listening gain is employed and listening restricted to the first 0.3 seconds.  High frequency dither noise then becomes quite noticeable.  [If this is not the case, your equipment is probably not playing back at 24-bits, but at only 16-bits.]

It is a much more difficult (impossible?) exercise to ABX at a moderate listening level.  Before doing any comprehensive listening tests at moderate levels, I would like to get the feedback of AndyH or others as to whether other dithering techniques are to be preferred.

If anyone is thinking of uploading an alternative dithered version, a 16-bit format would do.  I put my "dithered to 16 bits" version into a 24-bit format simply to reduce the likelihood of some software players processing it differently to the 24-bit original file when comparing playback.

At very high listening levels I find that the version that has been reduced to 16bits sounds a little duller to my ears.  However, my hearing becomes affected very rapidly at such volume levels, so that after a few repetitions I cannot hear any differences.

Short samples: 24 bit vs rendered to 16 bits dithered

Reply #1
Do you happen to have ABX logs available?

I can submit samples dithered with POW-r, UV22, Waves L1, L2, L3 and perhaps others if desirable. The two former algorithms retain 19 to 20 bits of resolution according to Bob Katz.

EDIT: Removed one portion after reading a few posts in the original thread.

Short samples: 24 bit vs rendered to 16 bits dithered

Reply #2
Do you happen to have ABX logs available?

I did not bother posting an ABX log as the exercise is dead easy when listening at high gain to the first 0.3 seconds.  I was able to get 10/10 (a 0.1% probability of guessing), without any trouble, and in a short time-frame.  But listening at a medium level would be much more challenging.

I can submit samples dithered with POW-r, UV22, Waves L1, L2, L3 and perhaps others if desirable. The two former algorithms retain 19 to 20 bits of resolution according to Bob Katz.

I have for a long time guessed that about 20 full actual bits (no dither) would be sufficient to satisfy the human ear, and in any event I understand that nominally 24-bit audio ADCs do not perform reliably for their least significant digits, or even if they did, real life microphones would generate too much noise to make the least significant digits meaningful.

I guess what would be desirable would be a dither that:
[blockquote](a)  generates less audible noise (fairly easy to establish by checking out the first 0.3 seconds of the 24-bit original at high gain); and/or,
(b) a dither that garners more of the missing detail, making the result sound closer to the original (probably quite difficult to test - perhaps more easily established by reducing the level of the original before further processing, i.e working with a lower bit depth version for preliminary testing purposes).[/blockquote]

I was not seeking the absolute best dither, merely either of the following:
[blockquote](1) an opinion that the dither I used (with the Audacity software) is close enough to optimal anyway and good enough to proceed with listening tests at lower volume (i.e tests not simply based on audible dither noise),
(2) to be provided with a dithered version that is of noticeably better quality and "up to scratch" to be used for further listening tests.[/blockquote]

Cheers.

Short samples: 24 bit vs rendered to 16 bits dithered

Reply #3
I downloaded the clips this morning but did not have time to do anything with them until this night. I’m writing this without reference to whatever may have been added since the thread was opened.

Listening only to the 0.4 seconds before the music starts, with the amplifier volume at maximum, I can easily hear a difference in the two files. I have enough experience to not be so foolish as to attempt to listen to the music at anywhere near that setting.

Comparing the 16 bit file I created from the original (in CoolEdit), on that opening 0.4 seconds, and the same volume setting, I hear no difference.

Still looking only to that opening bit before the music starts.
In all three files, CoolEdit’s  measurements (Analyze/Statistics) are quite near to each other. This is not the case if one generates silence at 32 bit, then dithers when converting to 16 bit. There the dither noise is very high compared with the original (just about the same measurements as after dithering this file’s noisy opening). I guess, in this file, the background noise being as high as it is accounts for most of the signal level; the dither noise adds relatively little.

The largest Statistic difference is in my 16 bit conversion: the peak amplitude is 2dB higher than the 24 bit file while the download 16 bit version is only 1dB higher. The noise shaping I used accounts for the difference in sound. It puts much more of the dither noise at frequencies higher than I can hear, while the downloaded 16 bit sample has more dither noise at lower frequencies.

I played with listening to the entire clip, as well as shorter sections of the music, comparing my 16 bit conversion to the 24 bit download. I could detect nothing upon which to make a guess about identity. Perhaps later I will have more energy to give to another attempt, but it has been a long day.

Again, I have no wish to read my posts in the thread which started this. I know that I have sometimes gone into more detail than saying “properly resampled and dithered” or whatever I wrote in the first post, but I don’t know if I was as specific somewhere in that thread as I have been in other posts on this topic.

Unless the file has no quiet passages, fade outs, or something of that aspect, it is not hard to make sure the dither is audible. I can dither so that the difference at the beginning of the downloaded clip is audible at a somewhat lower volume setting, but I don’t consider that relevant to the question. Can any difference in the music be distinguished? Can the between track sections even be heard at reasonable listening levels, let alone heard to be different?

I don’t see that between tracks is relevant anyway. While it is something that one hears in some recordings, if that, and only that, is somewhat different in sound from the original master, who is ever going to know it without access to that original master? One isn’t going to be able to say that the character of the between track sections somehow reduces the quality of the recording (unless it was so in the original recording, of course).

Fade in sometimes starts below the level of audibility and fade out sometimes goes down below audibility. That is not a part of the performance for which one is expected to crank the volume to maximum. The fade in and fade out, as such, are part of the performance. If one could hear a difference, while there was actually music signal present, with the volume setting remaining unchanged from the level that made the main part of the performance enjoyable, then there would be a legitimate difference.

Short samples: 24 bit vs rendered to 16 bits dithered

Reply #4
Thanks for doing this, MLXXX

I can't download the files (due to a restrictive firewall), but I would appreciate it if somebody would try apply Sebastian's noise shaped dither from the end of this thread, and try to ABX again. Be careful turning it up all the way, though - it could damage tweeters.

Short samples: 24 bit vs rendered to 16 bits dithered

Reply #5
[...] but I would appreciate it if somebody would try apply Sebastian's noise shaped dither from the end of this thread [...]


Code: [Select]
>java -jar requant.jar SprightlyHandelclip.wav SprightlyHandelclip_b16_d17_slameath48.wav 16 -d1.7 -s@lameath48.sos


Cheers,
SG

Short samples: 24 bit vs rendered to 16 bits dithered

Reply #6
Thanks very much SG,

To my ears, the short quiet section preceding the first notes of the piano sounds almost the same in your "chained 2nd-order sections in direct form II" dithered version as in the original clip.  Very hard to pick apart. (A fraction of the audible noise level introduced by the Audacity triangular dither I used.)

In comparing the two versions by ear I wouldn't say your dithered version is noisier.  It sounds about as noisy as the original 24-bit recording to me, just a subtle difference in the 'tone' of the noise.

So, it should be possible to do ABX tests with the gain at a fairly high setting and the background noise should sound the same (if it can be heard at all) and not be a tell-tale sign of dither.

[I don't know whether I'll be able to do listening tests tonight.  I have a cold.  I think listening for subtle differences is best done with a clear and alert mind.]

Short samples: 24 bit vs rendered to 16 bits dithered

Reply #7
Thanks SebastianG. I'll try to ABX when I got home this evening and report back.

Short samples: 24 bit vs rendered to 16 bits dithered

Reply #8
As I am playing with SebastianG's filter in lossyWAV and this thread pertains to bit-reduction, I tried SprightlyHandel with lossyWAV......

..... and it crashed with an FMT chunk error. Apparently, the posted WAV has 2 bytes appended to the FMT chunk. So, I amended lossyWAV to take this into account and successfully processed the WAV file (lossyWAV -3 -shaping 1.000, average bits removed: 6.97) and then encoded it with FLAC -5 - the result is attached.

[edit] Processed file removed - lossyWAV version used to process it was well out of date. [/edit]
lossyWAV -q X -a 4 -s h -A --feedback 2 --limit 15848 --scale 0.5 | FLAC -5 -e -p -b 512 -P=4096 -S- (having set foobar to output 24-bit PCM; scaling by 0.5 gives the ANS headroom to work)

Short samples: 24 bit vs rendered to 16 bits dithered

Reply #9
Has the 24-bit musical Grail been found?
Er no.  After many hours of web-surfing, file converting, and abx-ing, I have come up empty handed.

Although it goes against the grain to admit it, I cannot find any evidence that converting a 48KHz/24bit recording of music to 16 bits with a noise shaped dither makes any significant difference at realistic listening levels.

There is a logic in dithering.  It is effectively a supplementary sampling process albeit at a maximum of a little less than a half of the nominal sampling frequency.  (It is interesting that even using a 16-bit sound card for capturing, noise present in the electronics in the analogue recording chain could provide useful dither.)

Comparing a well dithered 16-bit rendered version with the original SprightlyHandel.wav
For the record, I can ABX the noise in SebastianG's 16-bit version if I listen very carefully at very high gain.  Here is a result I obtained earlier this evening when listening to the first 0.3 seconds of the file I uploaded about a week ago and the 16bit dithered version SebastianG kindly provided:

[blockquote]foo_abx 1.3.1 report
foobar2000 v0.9.5.1
2008/03/17 00:15:54

File A: D:\HA-uploadOrdownload\SprightlyHandelclip.wav
File B: D:\HA-uploadOrdownload\sprightlyhandelclip_b16_d17_slameath48.flac

00:15:54 : Test started.
00:18:05 : 01/01  50.0%
00:18:51 : 02/02  25.0%
00:22:23 : 03/03  12.5%
00:23:10 : 04/04  6.3%
00:23:36 : 05/05  3.1%
00:24:10 : 06/06  1.6%
00:24:55 : 07/07  0.8%
00:25:51 : Test finished.

----------
Total: 7/7 (0.8%)[/blockquote]

A secret with ABXing is not to answer unless you are sure.  I find my ears tire very quickly.  In the series above, there is quite a delay between answer two and answer three (over 3 minutes). That was because my ears were tired at that point and the samples were starting to sound the same.  However when I came back, my high frequency hearing had recovered, and I was able to hear a distinct difference.  (The dithered version has a higher pitched noise than the original.)

- MLXXX

P.S.  I would not say it is impossible to hear differences as between 16-bits dithered and a nominal 24-bits but the effect is so subtle as to be beyond what my ears can reliably report.  On a few occasions in my testing I thought I could hear the followng difference but after a short while my ears tired and I could not get a persuasive ABX report:
  • Some notes in the dithered version were slightly louder and raspier.
[/size]

As I am playing with SebastianG's filter in lossyWAV and this thread pertains to bit-reduction, I tried SprightlyHandel with lossyWAV......

..... and it crashed with an FMT chunk error. Apparently, the posted WAV has 2 bytes appended to the FMT chunk. So, I amended lossyWAV to take this into account and successfully processed the WAV file (lossyWAV -3 -shaping 1.000, average bits removed: 6.97) and then encoded it with FLAC -5 - the result is attached.


Hi NickC.  When I first played SprightlyHandel.lossy.flac, half an hour ago, it sounded definitely different, so I set it up on ABX and did two trials and was quite confident of my answers, and I found both answers were in fact correct.  However after that my ears, or concentration, tired and when I played it and SprightlyHandel.wav, they sounded the same! After a 10 minute break, I had another successful attempt but then they started sounding the same again.

(When I could hear the differences, the lossy version sounded a little less full, and slightly interrupted at times.)

Short samples: 24 bit vs rendered to 16 bits dithered

Reply #10
The more I look into dithering, the easier I find it to accept that the process if done optimally will not lead to audible distortion. 

As I do not appear to be able to upload files to the Scientific Forum I will upload a file  to here:
[attachment=4331:attachment]
It is a frequency distribution created by Cool Edit but there appears to be something wrong with the right channel that has led to some distortion.  This file is commented on at post #66 of Scientific/R&D Discussion > filtering, dither, and noiseshaping.

Short samples: 24 bit vs rendered to 16 bits dithered

Reply #11
The more I look into dithering, the easier I find it to accept that the process if done optimally will not lead to audible distortion.
The process if done optimally will not lead to ANY distortion, audible or otherwise. It will lead to an increase in the noise floor - but I think it's important, if pedantic, to make a distinction between THD and THD+N.

Short samples: 24 bit vs rendered to 16 bits dithered

Reply #12

The more I look into dithering, the easier I find it to accept that the process if done optimally will not lead to audible distortion.
The process if done optimally will not lead to ANY distortion, audible or otherwise. It will lead to an increase in the noise floor - but I think it's important, if pedantic, to make a distinction between THD and THD+N.

I think I might've done something wrong dithering the Handel sample to 16-bit.  There is CLEAR (at high volume) hiss at the very beginning.  I told Audacity to use "shaped" dithering.  I left it at 48 kHz.

Oh yeah, I'm using onboard audio (Analog Devices 1981HD in a ThinkPad T60).  I'm also using earbuds.  WTF?

Note:  I'm a total n00b at this stuff.

Short samples: 24 bit vs rendered to 16 bits dithered

Reply #13
wswartzendruber, noise shaped dithers vary.  The method Sebastian refers to at post #6 is quite effective (quieter than Audacity), and I have used it successfully myself with a number of files.

If you start with a very quiet 24-bit source and dither it to 16 bits, there will inevitably be slight additional noise or a slight change in the 'colour' of the noise after dithering, if you insist on setting the playback gain very high.  Having the gain that high, you will probably overload your speakers or your hearing when the playback reaches the loud parts of the recording.