IPB

Welcome Guest ( Log In | Register )

2 Pages V   1 2 >  
Reply to this topicStart new topic
The effects of musical genre upon ease/degree of compression, [split from “What music genres do you listen to?”, thread 96042]
pawelq
post Jul 17 2012, 16:09
Post #1





Group: Members
Posts: 541
Joined: 20-December 05
From: Springfield, VA
Member No.: 26522



QUOTE (IgorC @ Jul 17 2012, 00:06) *
Any audio codec has a "difficult" and "easy" genres to compress.



You are oversimplifying. "Classical music" is often considered a single genre*, yet the acoustic structure and consequently the ease/difficulty of compression vary enormously. For example, a harpsichord piece will be typically hard to compress due to very fast attacks and wide frequency range. A mellow and quiet full-orchestra piece, will be much easier. Respective examples from my collection (all FLAC, same quality) are up to 983 kbps for a two-harpsichord piece by Bach vs. 523 kbps for a slow movement of Beethoven's symphony or 418 kbps for a slow movement of Mozart's piano concerto

*which is misleading in the first place, as you can probably find much more musical and/or textural variety within output of a single classical composer, or sometimes within one piece, let alone the "classical" genre spanning ~500 years of history of music, than in entire "genres" of popular music.


--------------------
Ceterum censeo, there should be an "%is_stop_after_current%".
Go to the top of the page
+Quote Post
IgorC
post Jul 17 2012, 17:16
Post #2





Group: Members
Posts: 1582
Joined: 3-January 05
From: ARG/RUS
Member No.: 18803



This oversimplification is a first approximation. It's better than nothing.
If You will try some modern encoders You will see that there are general tendencies to increase bitrate on some particular genres.

In fact I see it much simpler. There are a few things like transients, tonality etc.
Generally classical music and Jazz contain more tonality while Rock and similar genres (rythmic genres) have more transients. Here oversimplification is completely valid.

This post has been edited by IgorC: Jul 17 2012, 17:29
Go to the top of the page
+Quote Post
pawelq
post Jul 17 2012, 17:50
Post #3





Group: Members
Posts: 541
Joined: 20-December 05
From: Springfield, VA
Member No.: 26522



QUOTE (IgorC @ Jul 17 2012, 12:16) *
If You will try some modern encoders You will see that there are general tendencies to increase bitrate on some particular genres.

Like I showed with my examples, bitrates may be hugely different for the same "genre" of classical.

QUOTE (IgorC @ Jul 17 2012, 12:16) *
In fact I see it much simpler. There are a few things like transients, tonality etc.

Agreed. My point is that mI predict that "genre" would have only at best a weak correlation with acoustic features that affect compression. Especially if you treat classical as a single genre.

QUOTE (IgorC @ Jul 17 2012, 12:16) *
Generally classical music and Jazz contain more tonality while Rock and similar genres (rythmic genres) have more transients. Here oversimplification is completely valid.

Your "generally" is the oversimplification that bothers me. I had a similar conversation before with people who claimed that heavy rock/heavy metal would be more difficult to compress than "classical" due to distortion, inharmonic content, etc. They claimed they could "heavy-metallize" my harpsichord sample by adding more distortion/noise/mistuning (the argument was in heavy metal you have two guitarists who play slightly differently) and that sample would require even higher bitrate. Thus supposedly "proving" that harpsichord is not as hard to compress as distorted heavy metal. They never came back with a such sample. Conversely, I was able to show that adding distortion and overlaying another slightly mistuned and time-shifted copy of the same harpsichord piece actually reduced bitrate for the same quality of Ogg Vorbis (160 kbps vs 195 kbps). Similarly, adding white noise led to bitrate reduction.

They ended up trying to use arguments from influence of noise in digital images on JPEG compression, which does not even deserve a comment.


--------------------
Ceterum censeo, there should be an "%is_stop_after_current%".
Go to the top of the page
+Quote Post
Porcus
post Jul 17 2012, 18:21
Post #4





Group: Members
Posts: 1995
Joined: 30-November 06
Member No.: 38207




The individual codecs' performances do indeed vary quite a lot over material. At http://www.hydrogenaudio.org/forums/index....st&p=800823 I compiled tests with my lowest-bitrate CD (an Edith Piaf compliation FLACing in at 251, where TAK -p4m WavPack High big time) and my highest-bitrate CD (a Merzbow album @ 1345 in FLAC, where WavPack High saves 96 kb/s over Tak -p4m).

Average over my non-classical FLACs (16-bits only, and bootlegs excluded) is 921 kb/s.




As for classical music: Following-up on pawelq, I sorted by bitrate my boxed set of Bach's collected works (6 days 19 minutes). All ripped to FLAC -8, average bitrate at 661.
Disregarding a half-minute vocal piece @ 375, the lowest bitrate found is Partitia for solo flute in A minor BWV 1013, averaging 409 over 18 minutes. That's allemande@425, corrente@420, sarabande@390 and bourrée@403 if anyone cares.
In the other end, there is a total of 69 minutes of keyboard music (probably all harpsichord, I didn't bother to check everything) where each track is 1000 or more. Up to 1048.

That's a pretty good variety, but a similar variety can be found on even a single album by jazz saxophonist John Zorn (both with Naked City and solo) and black metal turning whatever-they-want band Ulver.


--------------------
One day in the Year of the Fox came a time remembered well
Go to the top of the page
+Quote Post
Porcus
post Jul 17 2012, 19:04
Post #5





Group: Members
Posts: 1995
Joined: 30-November 06
Member No.: 38207



QUOTE (pawelq @ Jul 17 2012, 18:50) *
I had a similar conversation before with people who claimed that heavy rock/heavy metal would be more difficult to compress than "classical" due to distortion, inharmonic content, etc.


Since I'm at it:

14 percent of my (non-classical) CD tracks exceed in bitrate the single highest piece of classical music I own (which 1048 kb/s). Rammstein's “Mutter” and Scorpions' “Blackout” and “Savage Amusement” each exceed 1048 on average, to pick two well-known artists. So do the remastered version of Slayer's “Reign in Blood” and “South of Heaven” (but not the originals). I have a few albums where each and every track exceed the 1048 mark too, in case you are interested.

(I have about 19 days of classical music. Which means that Bach takes up 3/7 of it.)

This post has been edited by Porcus: Jul 17 2012, 19:06


--------------------
One day in the Year of the Fox came a time remembered well
Go to the top of the page
+Quote Post
pawelq
post Jul 17 2012, 19:38
Post #6





Group: Members
Posts: 541
Joined: 20-December 05
From: Springfield, VA
Member No.: 26522



QUOTE (Porcus @ Jul 17 2012, 14:04) *
So do the remastered version of Slayer's “Reign in Blood” and “South of Heaven” (but not the originals).


This is actually a very interesting observation, which shows that required bitrate can even depend on mastering not only within the same genre, but for the same tracks.

I am not trying to say that heavy metal cannot be difficult to compress, my point is that genre alone is probably a weak predictor of required bitrate.


I have these Slayer CDs somewhere (originals, I believe), I may later compare them to that harpsichord piece at the same FLAC compression level. Just out of curiosity


--------------------
Ceterum censeo, there should be an "%is_stop_after_current%".
Go to the top of the page
+Quote Post
Porcus
post Jul 17 2012, 20:40
Post #7





Group: Members
Posts: 1995
Joined: 30-November 06
Member No.: 38207



QUOTE (pawelq @ Jul 17 2012, 20:38) *
This is actually a very interesting observation, which shows that required bitrate can even depend on mastering not only within the same genre, but for the same tracks.


And for the purposes of attempting a «representative test corpus» ... well.


Now I would guess the explanation for the differences mentioned, would schematically be
Dynamic compression --> brings down peak --> turn up volume --> bring in stuff that was below the LSB (typically, more from the noise floor, which is harder to compress)

And some albums aren't even mastered at full peak in the original version, but these weren't too far from it.



ReplayGain values for three Slayer albums:
RiB: -3.8 vs -11.1. The former peaking at .98, whith one track peaking at .765.
SoH: -4.3 vs -10.9. The former peaking at 1, but one track down at .83. On the latter, every track is normalized to peak=.999969
Seasons: -6.66 (I kid you not) vs -10.9 -- again with every track normalized to that peak=.999969

(Since I was too careless upon ripping, I don't actually know for sure which one is remastered, but I would bet the loudness war victims are newer.)



--------------------
One day in the Year of the Fox came a time remembered well
Go to the top of the page
+Quote Post
pawelq
post Jul 18 2012, 00:41
Post #8





Group: Members
Posts: 541
Joined: 20-December 05
From: Springfield, VA
Member No.: 26522



QUOTE (Porcus @ Jul 17 2012, 15:40) *
Now I would guess the explanation for the differences mentioned, would schematically be
Dynamic compression --> brings down peak --> turn up volume --> bring in stuff that was below the LSB (typically, more from the noise floor, which is harder to compress)


Looks plausible, buy on the other hand, wouldn't dynamic compression make transients less "transienty"?


I don't use ReplayGain, but I scanned the three classical tracks that I mentioned before, and here are the results:

Two harpsichords (Bach, Contrapunctus IX from Kunst de Fuge, Koopman & Mathot), 983 kbps, track gain -1.90 dB, peak 0.788208
Symphony slow movement (Beethoven, Adagio molto from IX Symphony, Bernstein), 523 kbps. track gain +9.28 dB, peak 0.710114
Piano concerto slow movement (Mozart, Adagio from Piano Concerto No. 23, Zacharias & Zinman), 418 kbps, track gain +12.44 dB, peak 0.228149


--------------------
Ceterum censeo, there should be an "%is_stop_after_current%".
Go to the top of the page
+Quote Post
IgorC
post Jul 18 2012, 01:04
Post #9





Group: Members
Posts: 1582
Joined: 3-January 05
From: ARG/RUS
Member No.: 18803



QUOTE (pawelq @ Jul 17 2012, 13:50) *
QUOTE (IgorC @ Jul 17 2012, 12:16) *
If You will try some modern encoders You will see that there are general tendencies to increase bitrate on some particular genres.

Like I showed with my examples, bitrates may be hugely different for the same "genre" of classical.

The point is what doesn't happen on particular albums but what happens on average.

Or are you trying to say that if we randomly pick let say 30 albums of rock genre and 30 albums of classical genre and if one codec on avearge increases bitrates for rock
and decreases it for classical music there is any significant chance to get inverse results?
Is that what You are trying to say?

Ask any developer of audio codecs if tonality is the predominant for classic music and transients are predominant for rock music. We already know the answer.


QUOTE (pawelq @ Jul 17 2012, 13:50) *
QUOTE (IgorC @ Jul 17 2012, 12:16) *
Generally classical music and Jazz contain more tonality while Rock and similar genres (rythmic genres) have more transients. Here oversimplification is completely valid.

Your "generally" is the oversimplification that bothers me.

Would You still let me live in this world?


QUOTE (pawelq @ Jul 17 2012, 13:50) *
I had a similar conversation before with people who claimed that heavy rock/heavy metal would be more difficult to compress than "classical" due to distortion, inharmonic content, etc. They claimed they could "heavy-metallize" my harpsichord sample by adding more distortion/noise/mistuning (the argument was in heavy metal you have two guitarists who play slightly differently) and that sample would require even higher bitrate.

You and your oponents forget one and the most important variable: encoder or format. There are some encoders or formats that compress better tonality while other transients and noise like signals on average. (just in case if You will say : "But there are some except...")


QUOTE (pawelq @ Jul 17 2012, 13:50) *
Conversely, I was able to show that adding distortion and overlaying another slightly mistuned and time-shifted copy of the same harpsichord piece actually reduced bitrate for the same quality of Ogg Vorbis (160 kbps vs 195 kbps). Similarly, adding white noise led to bitrate reduction.

Did it involve ABX testing when You speak about same quality? Watch out, TOSX is prepared.

This post has been edited by IgorC: Jul 18 2012, 01:15
Go to the top of the page
+Quote Post
db1989
post Jul 18 2012, 07:38
Post #10





Group: Super Moderator
Posts: 5275
Joined: 23-June 06
Member No.: 32180



QUOTE (IgorC @ Jul 18 2012, 00:04) *
QUOTE (pawelq @ Jul 17 2012, 13:50) *
QUOTE (IgorC @ Jul 17 2012, 12:16) *
Generally classical music and Jazz contain more tonality while Rock and similar genres (rythmic genres) have more transients. Here oversimplification is completely valid.

Your "generally" is the oversimplification that bothers me.

Would You still let me live in this world?
What is this supposed to mean? I do not want to have to moderate a personal argument started from no basis. TOS #2 “is prepared”.

QUOTE (IgorC @ Jul 18 2012, 00:04) *
QUOTE (pawelq @ Jul 17 2012, 13:50) *
Conversely, I was able to show that adding distortion and overlaying another slightly mistuned and time-shifted copy of the same harpsichord piece actually reduced bitrate for the same quality of Ogg Vorbis (160 kbps vs 195 kbps). Similarly, adding white noise led to bitrate reduction.

Did it involve ABX testing when You speak about same quality? Watch out, TOSX is prepared.
It probably did. Besides, there is no need to seize every opportunity to challenge someone. Again, please do not say things that might provoke an argument for no reason.

The actual discussion about genres is proving to be interesting now that it has received some substantiation. If any more data can be gathered or if more defined conclusions can be reached, everyone will benefit. Why spoil that? Let’s keep things constructive.
Go to the top of the page
+Quote Post
pawelq
post Jul 18 2012, 17:25
Post #11





Group: Members
Posts: 541
Joined: 20-December 05
From: Springfield, VA
Member No.: 26522



QUOTE (IgorC @ Jul 17 2012, 20:04) *
The point is what doesn't happen on particular albums but what happens on average.

Or are you trying to say that if we randomly pick let say 30 albums of rock genre and 30 albums of classical genre and if one codec on avearge increases bitrates for rock
and decreases it for classical music there is any significant chance to get inverse results?
Is that what You are trying to say?



I am trying to say that the even if the means of distributions are different in the same direction as you predict, the standard deviation (or whatever measure of dispersion will you choose) will be so huge and the overlap of distribution so large that it will preclude any reliable precition as to the required bitrate based on the "genre" alone. Even if the difference will be significant in the sense of statistical hypothesis testing. Especially if you treat "classical music" with it's enormous span of musical factures, styles, numbers of instruments involved in a piece, as a single "genre".

I am not saying that you won't get a wide distribution also with rock or whatever, but being much more familiar with classical and having FLAC'ed many classical CDs, I can be more sure about classical. Even within one classical piece/album - same composer, same musical composition, same performers, same mastering - I can find differences in FLAC birate between tracks to be on the order of 300 kbps (760 vs. 457 in an example I am looking at now).



QUOTE (IgorC @ Jul 17 2012, 20:04) *
Ask any developer of audio codecs if tonality is the predominant for classic music and transients are predominant for rock music. We already know the answer.


Do we? I predict that again you will have a lot of overlap. The harspichord that I keep mentioning is a very "transienty" instrument. Piano has very definite transient onsets as well. Or have you ever heard Varese's "Ionisation"? It's a "classical" piece, and check out it's instrumentation http://en.wikipedia.org/wiki/Ionisation_(Var%C3%A8se) . Do you think it's very tonal?

I made masurement of harmonicity (ratio of harmonic to inharmonic components) of four tracks with audio analysis program Praat. The sample is small for now but calculation takes quite a lot of time, and also the results are probably not 100% accurate because I had the rock pieces only in VBR mp3. So it's just for a start.


Ionisation (Boulez): mean hamonicity -0.82 dB; maximum harmonicity 47.36 dB
Slayer, War ensemble: 0.86 dB; 35.52 dB
Marylin Manson, I don't like drugs: 0.72 dB; 48.82 dB
Contrapunctus IX, Koopman/Mathot: 2.01 dB; 34.17 dB


QUOTE (IgorC @ Jul 17 2012, 20:04) *
You and your oponents forget one and the most important variable: encoder or format. There are some encoders or formats that compress better tonality while other transients and noise like signals on average.
This is probably true. But how much does it help if "genre" is, as I argue, a rather poor predictor of transientiness or noisiness of the acoustic signal?


QUOTE (IgorC @ Jul 17 2012, 20:04) *
Did it involve ABX testing when You speak about same quality? Watch out, TOSX is prepared.

No, it involved the same Vorbis quality setting. I compared bitrate resulting from compression of a tracks (i.e., harpsichord vs. distorted harpsichord, or harspichord vs. harpsichord+noise). ABX not applicable, I think.


--------------------
Ceterum censeo, there should be an "%is_stop_after_current%".
Go to the top of the page
+Quote Post
IgorC
post Jul 18 2012, 20:59
Post #12





Group: Members
Posts: 1582
Joined: 3-January 05
From: ARG/RUS
Member No.: 18803



QUOTE (pawelq @ Jul 18 2012, 13:25) *
QUOTE (IgorC @ Jul 17 2012, 20:04) *
Did it involve ABX testing when You speak about same quality? Watch out, TOSX is prepared.

No, it involved the same Vorbis quality setting. I compared bitrate resulting from compression of a tracks (i.e., harpsichord vs. distorted harpsichord, or harspichord vs. harpsichord+noise). ABX not applicable, I think.

it's not guarantee that the quality maintans constant for the same Q Vorbis setting.

QUOTE (pawelq @ Jul 18 2012, 13:25) *
ABX not applicable, I think.

You are on wrong forum.


Sorry but all your critics have no nothing in common what it's happening in the world of lossy encoders.
The simple treatment of classic and rock as music genres which contains mostly tonality and transient stuff works very well. The proof of that is Opus encoder with its new VBR. And it makes an excelent job.
While You keep talking about that there is an important amount of transients in classic genre there is no even sense to talk about that because it's per frame decision.
It's not like classic music will be all compressed in tonality mode on all song.

All good encoder needs is a good calibration for VBR and good decision per frame and frequency band. No need a description of quantum effect, even less what happens during on particular moments because lossy encoders has a good time resolution ( down to 2.5-3 ms ) as well some lossy encoders can handle transients and tonality at the same time at different frequencies bands.

This post has been edited by IgorC: Jul 18 2012, 21:22
Go to the top of the page
+Quote Post
Porcus
post Jul 18 2012, 22:27
Post #13





Group: Members
Posts: 1995
Joined: 30-November 06
Member No.: 38207



QUOTE (IgorC @ Jul 18 2012, 21:59) *
QUOTE (pawelq @ Jul 18 2012, 13:25) *
ABX not applicable, I think.

You are on wrong forum.


Depends. Just because someone (with a nick indicating a different mother tongue than English) every now and then mixes up the wording over quality settings vs quality target vs subjective audio quality, I wouldn't suggest (s)he should GTFO from here.


But this thread is getting a bit too noisy. And, may I add, that could be because it is not very clear what was behind the poll questions.
Is the purpose is to get a rough idea of what a representative test corpus for encoders? Your posting #8 might indicate so, but the fact that it is posting #8 and not the start of the thread, might suggest otherwise. Still I don't understand whether this 'occational' means I should tick one or six boxes.


--------------------
One day in the Year of the Fox came a time remembered well
Go to the top of the page
+Quote Post
pawelq
post Jul 18 2012, 22:45
Post #14





Group: Members
Posts: 541
Joined: 20-December 05
From: Springfield, VA
Member No.: 26522



QUOTE (IgorC @ Jul 18 2012, 15:59) *
it's not guarantee that the quality maintans constant for the same Q Vorbis setting.

Obviously, but it's the best approximation.


QUOTE (IgorC @ Jul 18 2012, 15:59) *
QUOTE (pawelq @ Jul 18 2012, 13:25) *
ABX not applicable, I think.
You are on wrong forum.
I am afraid that you did not understand. I compared compression ratios (bitrates) of A. original harpsichord music vs. B. same harpsichord music but with distortion and other processing added, as well as A. original harpsichord music vs. C. same harpsichord music but with noise added. Can you explain how ABX testing would help me decide whether lossy-compressed A is of same quality as lossy-compressed B or C if uncompressed A was different from uncompressed B (and C) in the first place?


QUOTE (IgorC @ Jul 18 2012, 15:59) *
The simple treatment of classic and rock as music genres which contains mostly tonality and transient stuff works very well. The proof of that is Opus encoder with its new VBR. And it makes an excelent job.

Did I say something that opposes what you just said? Yes, both classical and rock music contain tonal/harmonic contents as well as transients. And these affect compression ratio. And I understand (maybe incorrectly, but I don't know much about Opus, please correct me if I am wrong) that Opus is particularly good at taking advantage of this distinction by applying different processing to harmonic vs. non-harmonic pieces of audio.

My point is that the genre label is not a reliable predictor of harmonic/tonal vs. non-harmonic content in a file, and therefore can be misleading if used as a basis for using this or that processing/compression approach within an encoder.



QUOTE (IgorC @ Jul 18 2012, 15:59) *
While You keep talking about that there is an important amount of transients in classic genre there is no even sense to talk about that because it's per frame decision.
It's not like classic music will be all compressed in tonality mode on all song.

Exactly! The decision is based on the actual audio content of each frame, not on the genre label of the track! This is why I am opposing your justification of the poll, which was "Bitrate varies considerably per genre.
The results of this poll can be useful for bitrate verifications during preparation/discussion of public tests as well as for developers (calibration of VBR modes, optimization of compression etc.)
"

They can't be useful, because genre does not predict well nhow compression should be optimized for a particular track or frame (not to mention that I have seen tracks with genre tag completely wrong).




QUOTE (IgorC @ Jul 18 2012, 15:59) *
All good encoder needs is a good calibration for VBR and good decision per frame and frequency band. No need a description of quantum effect, even less what happens during on particular moments because lossy encoders has a good time resolution ( down to 2.5-3 ms ) as well some lossy encoders can handle transients and tonality at the same time at different frequencies bands.

Exactly. A good encoder needs what you just said above. It does not need to know whether the track is labeled as Classical or New-Metal or whatever.


--------------------
Ceterum censeo, there should be an "%is_stop_after_current%".
Go to the top of the page
+Quote Post
IgorC
post Jul 19 2012, 04:12
Post #15





Group: Members
Posts: 1582
Joined: 3-January 05
From: ARG/RUS
Member No.: 18803



QUOTE (pawelq @ Jul 18 2012, 18:45) *
QUOTE (IgorC @ Jul 18 2012, 15:59) *
QUOTE (pawelq @ Jul 18 2012, 13:25) *
ABX not applicable, I think.
You are on wrong forum.
I am afraid that you did not understand. I compared compression ratios (bitrates) of A. original harpsichord music vs. B. same harpsichord music but with distortion and other processing added, as well as A. original harpsichord music vs. C. same harpsichord music but with noise added. Can you explain how ABX testing would help me decide whether lossy-compressed A is of same quality as lossy-compressed B or C if uncompressed A was different from uncompressed B (and C) in the first place?

I've undrestood You previously (see observation down).
Your statements:

QUOTE (pawelq @ Jul 17 2012, 13:50) *
Conversely, I was able to show that adding distortion and overlaying another slightly mistuned and time-shifted copy of the same harpsichord piece actually reduced bitrate for the same quality of Ogg Vorbis (160 kbps vs 195 kbps).


QUOTE (pawelq @ Jul 18 2012, 13:25) *
ABX not applicable, I think.


So do You intentionally change the source and then claim the same quality? On top of that You think (???) that ABX isn't apllicable?
The same quality of what? Original source or filtered? Or is it just the same?

Sorry, I'm trying to understand your logic.
Go to the top of the page
+Quote Post
IgorC
post Jul 19 2012, 04:16
Post #16





Group: Members
Posts: 1582
Joined: 3-January 05
From: ARG/RUS
Member No.: 18803



QUOTE (pawelq @ Jul 18 2012, 18:45) *
My point is that the genre label is not a reliable predictor of harmonic/tonal vs. non-harmonic content in a file, and therefore can be misleading if used as a basis for using this or that processing/compression approach within an encoder.

Nobody was going to put any genre label (or whatever it is) on file basis.
I don't know why You even discuss it.


QUOTE (pawelq @ Jul 18 2012, 13:25) *
I am trying to say that the even if the means of distributions are different in the same direction as you predict, the standard deviation (or whatever measure of dispersion will you choose) will be so huge and the overlap of distribution so large that it will preclude any reliable precition as to the required bitrate based on the "genre" alone. Even if the difference will be significant in the sense of statistical hypothesis testing. Especially if you treat "classical music" with it's enormous span of musical factures, styles, numbers of instruments involved in a piece, as a single "genre".

Agree.
Nobody says it will be 90-95% predictable bitrate for lossy encoder.

Do users require a final bitrates be within +/- 5-10% from the average per track or even per album? wink.gif


This post has been edited by IgorC: Jul 19 2012, 04:48
Go to the top of the page
+Quote Post
greynol
post Jul 19 2012, 05:00
Post #17





Group: Super Moderator
Posts: 10348
Joined: 1-April 04
From: San Francisco
Member No.: 13167



There is no need for ABX here unless you wish to demonstrate that the original and the distorted versions of harpsichord giving rise to different bitrates using the same configured quality setting sound different.

I believe there is some merit to the notion that different genres require different bitrates (lossy or lossless) on average and that for classical music, harpsichord is a well known exception.

Whether someone may be able to make good use of genres in order to make better/more efficient coding decisions is interesting but isn't something I've put much thought into personally. My initial response is that won't matter much, but I've been accused of being a pessimist on more than one occasion. wink.gif

This post has been edited by greynol: Jul 19 2012, 05:10


--------------------
Breath is found in plots and DR figures.
Go to the top of the page
+Quote Post
IgorC
post Jul 19 2012, 07:46
Post #18





Group: Members
Posts: 1582
Joined: 3-January 05
From: ARG/RUS
Member No.: 18803



QUOTE (pawelq @ Jul 17 2012, 13:50) *
Conversely, I was able to show that adding distortion and overlaying another slightly mistuned and time-shifted copy of the same harpsichord piece actually reduced bitrate for the same quality of Ogg Vorbis (160 kbps vs 195 kbps). Similarly, adding white noise led to bitrate reduction.

It will be great if You will open a new topic and share details of your findings (application, results, conclusion, analysis etc.). I should calm myself too but I beleive such rather big claim should be verified.


Go to the top of the page
+Quote Post
Porcus
post Jul 19 2012, 11:46
Post #19





Group: Members
Posts: 1995
Joined: 30-November 06
Member No.: 38207



QUOTE (greynol @ Jul 19 2012, 06:00) *
Whether someone may be able to make good use of genres in order to make better/more efficient coding decisions is interesting but isn't something I've put much thought into personally. My initial response is that won't matter much, but I've been accused of being a pessimist on more than one occasion. wink.gif


I would suppose that testing by genres, subgenres, and even mastering ideals (peak-normalized or not, loudness-war or not) could be tools for improving both encoding algorithms and codecs themselves – certainly for maximum feasible compression, but presumably also for speed-given-ratio. But when that numbercrunching procedure is done, I would be very surprised if there would be much to gain compression-wise by spending one or two bits on an initial genre-flag. And speed-wise? Listening to an entire track to guess genre first, would you get that paid back at a given compression level? I'd bet against.

I suppose from the result I reported http://www.hydrogenaudio.org/forums/index....st&p=800823 that TBeck doesn't have too much Japanoise in his collection, and likewise that Bryant could improve on the Edit Piaf thing had he taken the effort. (Would non-statisticians be likely to misinterpret the claim that “Merzbow appears to be out-of-sample”?) If these two CDs are/were representative of the respective codecs'/encoders' behaviour, then the choice of test corpus would have tremendous impact on the measured (relative!) performance.


--------------------
One day in the Year of the Fox came a time remembered well
Go to the top of the page
+Quote Post
Porcus
post Jul 19 2012, 12:31
Post #20





Group: Members
Posts: 1995
Joined: 30-November 06
Member No.: 38207



QUOTE (IgorC @ Jul 19 2012, 08:46) *
QUOTE (pawelq @ Jul 17 2012, 13:50) *
Conversely, I was able to show that adding distortion and overlaying another slightly mistuned and time-shifted copy of the same harpsichord piece actually reduced bitrate for the same quality of Ogg Vorbis (160 kbps vs 195 kbps). Similarly, adding white noise led to bitrate reduction.

It will be great if You will open a new topic and share details of your findings (application, results, conclusion, analysis etc.). I should calm myself too but I beleive such rather big claim should be verified.


While it looks strange at first glance, I don't think the claim describes any unwanted behaviour for lossy codecs -- indeed, I think the fallacy of the argument is that a lossy was used to compare.


Consider the following procedure:
(1) This is pretty much like noise.
(2) Psychoacoustic model says this is not very useful to the listener.
(3) We need to discard 80 percent of the signal anyway, so we discard this noise-alike thing.

Reasonable? Certainly. I recently learned that certain codecs (AAC and MPC) even use it actively: rather than discarding it, which would be at the risk of tilting the tonal balance, they substitute noise with noise-alike signals which are easily compressable. Perceptual noise substitution, it is called.

So what would be the effect of adding noise to harpsichord if the above procedure works well?
Without noise, this is a complex signal. A VBR encoder will increase bitrate.
With noise, this is a noisy signal. If there are overtones up there, they will be buried in noise, hence become useless, hence can be discarded. VBR encoder doesn't anymore see the need for high bitrate.

If the detection of 'noise-alikeness' is bad, then the algorithm could fail. It could mistake useless signal parts for important and spend bitrate on encoding them. It could mistake useful signal parts for noise and discard them. Certainly, the psychoacoustic models have improved over the years. Still I'd say it is risky to compare different samples (like, different genres) -- the encoder isn't necessarily equally good on both. And that isn't really necessary either, as long as you are encoding at a bitrate that is anyway transparent most of the time.

This post has been edited by Porcus: Jul 19 2012, 12:36


--------------------
One day in the Year of the Fox came a time remembered well
Go to the top of the page
+Quote Post
db1989
post Jul 19 2012, 12:46
Post #21





Group: Super Moderator
Posts: 5275
Joined: 23-June 06
Member No.: 32180



QUOTE (IgorC @ Jul 19 2012, 07:46) *
QUOTE (pawelq @ Jul 17 2012, 13:50) *
Conversely, I was able to show that adding distortion and overlaying another slightly mistuned and time-shifted copy of the same harpsichord piece actually reduced bitrate for the same quality of Ogg Vorbis (160 kbps vs 195 kbps). Similarly, adding white noise led to bitrate reduction.
It will be great if You will open a new topic and share details of your findings (application, results, conclusion, analysis etc.). I should calm myself too but I beleive such rather big claim should be verified.

Are people now required to submit a comprehensive set of data for every simple numerical claim they make? I must have missed that addendum to the ToS. But in seriousness, mentioning a difference in bitrate is not the same as making a statement about sound quality, for various reasons including relative ease of replicability – so please do not present it as though it is.

Also, btw, the only personal pronoun in English that receives a capital letter is I, not you.
Go to the top of the page
+Quote Post
Porcus
post Jul 19 2012, 13:22
Post #22





Group: Members
Posts: 1995
Joined: 30-November 06
Member No.: 38207



QUOTE (db1989 @ Jul 19 2012, 13:46) *
btw, the only personal pronoun in English that receives a capital letter is I, not you.


Well ... that might depend on whether one refers to (or, like presumably in this particular case, addresses) his local $DEITY laugh.gif


--------------------
One day in the Year of the Fox came a time remembered well
Go to the top of the page
+Quote Post
greynol
post Jul 19 2012, 14:20
Post #23





Group: Super Moderator
Posts: 10348
Joined: 1-April 04
From: San Francisco
Member No.: 13167



Can we use that to excuse errors in subject-verb agreement?
http://www.hydrogenaudio.org/forums/index....&pid=802685

If so I'll see if there's room to amend my sig. wink.gif


--------------------
Breath is found in plots and DR figures.
Go to the top of the page
+Quote Post
IgorC
post Jul 19 2012, 16:41
Post #24





Group: Members
Posts: 1582
Joined: 3-January 05
From: ARG/RUS
Member No.: 18803



QUOTE (pawelq @ Jul 17 2012, 12:09) *
You are oversimplifying.

At this point it's clear that we've talked about different things and You have accused me of oversimplification without getting into details.
It can lead only to misunderstanding.

This post has been edited by IgorC: Jul 19 2012, 16:43
Go to the top of the page
+Quote Post
pawelq
post Jul 20 2012, 04:51
Post #25





Group: Members
Posts: 541
Joined: 20-December 05
From: Springfield, VA
Member No.: 26522



QUOTE (greynol @ Jul 19 2012, 00:00) *
There is no need for ABX here unless you wish to demonstrate that the original and the distorted versions of harpsichord giving rise to different bitrates using the same configured quality setting sound different.
I only wanted to show that they enode to different bitrates at the same q. I even don't know how ABX could be used meaningfully in the context of that test. Ideally, I would prefer to show that the encoded version are degraded by lossy compression to the same degree, but I can't see how to do it even if I used very aggresive compression that would produce audible degradation - because the originals were different,.


QUOTE (greynol @ Jul 19 2012, 00:00) *
I believe there is some merit to the notion that different genres require different bitrates (lossy or lossless) on average and that for classical music, harpsichord is a well known exception.
I started doing some tests and looks like you and IgorC are much more right than I originally thought, at least for lossless. At first glance it looks like you might be not so much right about lossy, but I am going to do some more tests and I'll definitely come back to the thread with results. By the way, is there a way to quickly export/copy specific information about all files in a foobar2000 playlist? I would like to export name Tab code Tab bitrate Tab ReplayGain.


QUOTE (greynol @ Jul 19 2012, 00:00) *
Whether someone may be able to make good use of genres in order to make better/more efficient coding decisions
Again, I don't understand how using genre information in addition to (or instead of?) actual audio content could improve coding efficiency.

QUOTE (Porcus @ Jul 19 2012, 07:31) *
While it looks strange at first glance, I don't think the claim describes any unwanted behaviour for lossy codecs -- indeed, I think the fallacy of the argument is that a lossy was used to compare.


Consider the following procedure:
(1) This is pretty much like noise.
(2) Psychoacoustic model says this is not very useful to the listener.
(3) We need to discard 80 percent of the signal anyway, so we discard this noise-alike thing.

Reasonable? Certainly. I recently learned that certain codecs (AAC and MPC) even use it actively: rather than discarding it, which would be at the risk of tilting the tonal balance, they substitute noise with noise-alike signals which are easily compressable. Perceptual noise substitution, it is called.

So what would be the effect of adding noise to harpsichord if the above procedure works well?
Without noise, this is a complex signal. A VBR encoder will increase bitrate.
With noise, this is a noisy signal. If there are overtones up there, they will be buried in noise, hence become useless, hence can be discarded. VBR encoder doesn't anymore see the need for high bitrate.
That would be my explanation as well. By the way, the discussion that prompted me to do those tests was about lossy coding. The other person maintained that metal is more difficult to lossy-compress than classical because of being less harmonic, more noisy, more distorted, and more "mistuned". And he said that he could "metalicize" the harpsichord sample and make it even harder to compress. I showed that adding noise or distortion etc actually makes it easier to compress (when mesaured with bitrate at the same q setting). I think that in the case of lossy compression it is more about transients and much less about noise/distortion/inharmonicity.


QUOTE (IgorC @ Jul 19 2012, 11:41) *
At this point it's clear that we've talked about different things

I cannot exclude this possibility. So, again, what exactly the rationale of your genre poll in the context of improving audio encoders? And, by the way, does the item named "Classic" in your poll refer to so called "Classical music", i.e., to a certain tradition of Western educated music composed during last few centuries? Or to something else?


--------------------
Ceterum censeo, there should be an "%is_stop_after_current%".
Go to the top of the page
+Quote Post

2 Pages V   1 2 >
Reply to this topicStart new topic
1 User(s) are reading this topic (1 Guests and 0 Anonymous Users)
0 Members:

 



RSS Lo-Fi Version Time is now: 29th December 2014 - 04:32