Looking at the recent tests on 64kb files, I was wondering how come AAC underperforms compared to OGG, WMA and whatever else? I thought AAC was able to handle low bitrates very well due to it's type of encoding... if so, how come these other formats are beating it? Is it something that can be worked on in future releases from Psytel? Spose they would of been working mainly on high bitrate quality to whoop MPC's ass? :o)
Originally posted by Mac
I thought AAC was able to handle low bitrates very well due to it's type of encoding... if so, how come these other formats are beating it?
There isn't any high quality AAC implementation available that uses Intendity Stereo (although IS is part of the specs). That's the reason.
AAC+ (AAC+SBR) should be available soon and will address this issue as well.
Is it something that can be worked on in future releases from Psytel? Spose they would of been working mainly on high bitrate quality to whoop MPC's ass? )
High bitrate work on Psytel AACenc is nearly done, I would believe.
Ivan is probably working on low bitrate tuning more.
Regards;
Roberto.
Aah cool ) I heard something about AAC+SBR on the Psytel website, but didn't quite understand it!
I take it there's no reason to wait for this new version before encoding my wav files at -extreme? I won't gain any extra quality by waiting?
Originally posted by rjamorim
There isn't any high quality AAC implementation available that uses Intendity Stereo (although IS is part of the specs). That's the reason.
AAC+ (AAC+SBR) should be available soon and will address this issue as well.
High bitrate work on Psytel AACenc is nearly done, I would believe.
Ivan is probably working on low bitrate tuning more.
Regards;
Roberto.
Castanets encoding with different encoder at 128 kbps. My ratings:
1. Envivo AAC 4.9 (nearly perfect)
2. FhG AAC 5.0-0.2 (most the time better than Envivo, but strange loud clicks in the right channel von time to time)
3. HHI AAC 4.6 (some strange noise)
3. Psytel AAC 4.6 (some strange noise)
5. NEC 4.3 (one castanets was completely wrong, strange noise)
6. Lame 3.91 3.8
7. MPC 1.7.9 3.5 (muffled, strange noise)
8. ISO AAC 3.2 (some strange noise, castanets messed up)
9. tooLame 2.0 (worse noises, problems with the castagnets)
10. Philips AAC 1.0 (weird sound at all)
Hmmm, where do you get all these AAC encoders from? Or is it a case of you don't get them. If not, Ivan needs to make his the best just to teach them a lesson! )
Originally posted by Mac
Hmmm, where do you get all these AAC encoders from?
Just what I was wondering!
And I wonder how can Philips sound worse than ISO!
If not, Ivan needs to make his the best just to teach them a lesson! )
Well, keep in mind that, while castanets is a good sample to tune encoders, it isn't proper to compare encoders' quality using it. It's an exception sample, because it's not representative of generic music.
For your former question:
You can encode with -extreme safely now. Heh, I have files encoded with AACenc 1.6 -extreme that still sound perfect.
Regards;
Roberto.
PS: Please avoid using : o ) as smiley.
Or check "Disable smilies in This Post" before posting.
Originally posted by rjamorim
You can encode with -extreme safely now. Heh, I have files encoded with AACenc 1.6 -extreme that still sound perfect.
Really? How do you keep your files from sounding worse when new encoder versions come out? I don't know what I'm going to do when Vorbis 1.01 comes out.
Sorry, I just couldn't pass it up.
Originally posted by Ardax
Really? How do you keep your files from sounding worse when new encoder versions come out? I don't know what I'm going to do when Vorbis 1.01 comes out.
Well, they already sounded very good encoded with 1.6. If they sound OK to my ears, I see no reason to reencode.
Originally posted by Mac
Hmmm, where do you get all these AAC encoders from?
Or better yet, where can "I" get these encoders from?
Originally posted by rjamorim
Well, keep in mind that, while castanets is a good sample to tune encoders, it isn't proper to compare encoders' quality using it. It's an exception sample, because it's not representative of generic music.
One also has to consider what "generic" music truly is. Is my IDM generic music? Is my metal? What about my classical guitar music (filled with transients) like Bernd Steidl?
How do you define generic music? Is it what's on MTV? Is it the latest, greatest pop band? Classic rock?
For the most part, I can accept people saying things are not representative when talking about some admittedly obscure electronic music which a very large majority of the population wouldn't even consider music in the first place, but something like castanets surely doesn't fall in that category. A lot of music has rapid percussive sounds in it, and so a lot of music has the potential to be just "exceptional" as castanets.
Bottom line is how do you decide which samples can be used to compare codecs? It's kind of paradoxical to use samples which all encoders will sound equal on because they are extremely easy to encode. By the same token, any randomly chosen sample which ends up causing particular problems for an encoder (or many) can be said to be non-representative by just about anyone... so how do you decide where to draw the line?
Bit OT, yes, but I just wanted to add a bit of insight here
Originally posted by Dibrom
A lot of music has rapid percussive sounds in it
Yeah, but not as in castanets. Very few other music has such rapid percussion - and Castanets in nearly only percussion, too.
BESIDES, it's a well know fact that subband coders (MPC) handle these kinds of samples much better than transfom coders (Vorbis, MP3, AAC). So, based on these results, would you assume that MP3, AAC and Vorbis suck?
SO, my point is that, if you want to compare codecs quality, you can't use ONLY castanets as Mr. Klemm used (Although using it to compare encoders of a same format is relatively fair. Relatively, because one encoder can perform very well on every aspect of audio encoding, but wasn't properly tuned to handle transients). It's VERY well known that castanets is unfavorable to transform codecs.
If you want to test codecs quality, it's OK (i think) to include castanets. But you should be sure to include classical, pop, soft, metal, unplugged... as well. Both easy and hard samples. THAT would be a generic representative, IMO.
And, again, not ONLY castanets.
Hope I made myself clear.
Regards;
Roberto.
Originally posted by rjamorim
Yeah, but not as in castanets. Very few other music has such rapid percussion - and Castanets in nearly only percussion, too.
Hrmm.. this is what I'm not so sure about. I listen to a lot of music with similar sounds in it. How many other people do as well? I'm just saying... how can you quantify this exactly?
BESIDES, it's a well know fact that subband coders (MPC) handle these kinds of samples much better than transfom coders (Vorbis, MP3, AAC). So, based on these results, would you assume that MP3, AAC and Vorbis suck?
No, not at all. But I would believe it would be valid to make a statement about quality at least in a given situation if results hold true over a variety of similar samples, even if these samples may not be considered by any given person to be representative of generic music.
SO, my point is that, if you want to compare codecs quality, you can't use ONLY castanets as Mr. Klemm used (Although using it to compare encoders of a same format is relatively fair. Relatively, because one encoder can perform very well on every aspect of audio encoding, but wasn't properly tuned to handle transients). It's VERY well known that castanets is unfavorable to transform codecs.
I very much agree with this. 1 sample is surely nowhere enough. What I'm getting at though is what happens when someone wants to makes a comparison based on a slew of difficult samples. Say 20 or 30 of the most well known difficult to encode samples. Can a general statement about quality be made based on these? I think it can. That's at least how I initially did it with --alt-preset vs --r3mix. That's how I tuned things and how I came to the conclusion that the --alt-presets are better. If you can let someone say that 1 sample is not representative though (or worse, not a proper comparison), what's to stop them from saying all 30 are not representative? See what I mean?
If you want to test codecs quality, it's OK (i think) to include castanets. But you should be sure to include classical, pop, soft, metal, unplugged... as well. Both easy and hard samples. THAT would be a generic representative, IMO.
And, again, not ONLY castanets.
Hope I made myself clear.
Certainly I'm just wary of making a statement such as "it's not valid", because I think it's a very hard line to draw, and if you begin to make such statements in one situation (you or anyone else, nobody in particular), then it can lead to forming an ideology that critical samples (where the largest differences exist between codecs, IMO the more important situations) are no good for comparison at all.
Originally posted by Dibrom
Say 20 or 30 of the most well known difficult to encode samples. Can a general statement about quality be made based on these?
I think it probably would be valid.
But, instead choosing worst-case samples, I would rather choose samples of different music styles. (If these worst-case samples are "representative" - perfect!)
My rationalization:
If you use samples from different music stiles to tune your encoder, you'll be tuning it for the general public that will be using your encoder.
If you use only problem case samples, you'll be tuning your encoder for listening tests that use only these samples. You'll be tuning it for a limited part of the public that enjoys that kind of music too, but a much wider public would be reached using "generic samples"
After all, what's worth if an encoder performs superbly on Castanets and Fatboy, but fails on Pink Floyd? Such an encoder would be worthless!
Regards;
Roberto.
BTW: Just to clarify:
When I say "encoder", it's the implementation: Lame, MP3enc, AACenc...
When I say "codec", it's the format: MP3, AAC...
Originally posted by rjamorim
I think it probably would be valid.
But, instead choosing worst-case samples, I would rather choose samples of different music styles. (If these worst-case samples are "representative" - perfect!)
My rationalization:
If you use samples from different music stiles to tune your encoder, you'll be tuning it for the general public that will be using your encoder.
If you use only problem case samples, you'll be tuning your encoder for listening tests that use only these samples. You'll be tuning it for a limited part of the public that enjoys that kind of music too, but a much wider public would be reached using "generic samples"
Well the idea would of course be that these difficult samples
would come from many different genres. In fact, most of the time they do. Certainly the ones used to tune the --alt-preset's did.
I guess to further clarify what I'm getting at... if a person says or believes that each of these samples on their own are not representative or good for a valid comparison, what's to prevent someone from saying the exact same thing even if they are grouped up.
"They are all extreme cases.. so who cares?", etc.
After all, what's worth if an encoder performs superbly on Castanets and Fatboy, but fails on Pink Floyd? Such an encoder would be worthless!
lol!
After all, what's worth if an encoder performs superbly on Castanets and Fatboy, but fails on Pink Floyd? Such an encoder would be worthless!
All in all if you ask me it's just "another brick in the wall."
Funny that this topic of sample selection should come up now because I've been beating my head against the wall trying to think of a way to practically, but fairly choose samples, which of course will be much harder in a high bitrate test.
In the extreme (but fairest case), one would just choose randomly from a variety of genres without regard to how "difficult" they are to any of the encoders involved in a comparison. At high bitrates, though, one would probably have to choose many samples to find those difficult enough to discriminate between codecs. Ideally one would like to have (for the sake of argument) a dozen of these difficult samples to say something substantive about relative codec quality. But if only 10 percent (again for the sake of argument) of the randomly selected samples are difficult, there is a little problem -- one would have to go through 120 samples to find the difficult dozen.
120 samples to rate is clearly ridiculous. How does one fairly come up with difficult samples to use in a high-bitrate listening test?
ff123
120 samples to rate is clearly ridiculous. How does one fairly come up with difficult samples to use in a high-bitrate listening test?
I imagine that there is a fairly strong positive correlation between a 'challenging sample' for a particular codec, and the bitrate of the encoded file produced by that codec - assuming that you are using a quality based encode (--alt-preset standard, Vorbis -q 6, to name the two which I'm most interested in).
So you should be able to let the codecs themselves do some prescreening for you - if every/almost every quality based encoder uses less than the 'nominal' bitrate to encode that file, then just assume that the file is 'easy', and will not be useful as a discriminator in a listening test.
You are then left with a (hopefully smaller) collection of files which at least some of the encoders consider to be challenging. If not all of the encoders agree, then that's even better (I imagine there will be disagreements between the transform and subband encoders on some samples, which is just what you want).
Of course, you're relying on the encoders being relatively well tuned, and it could be argued that this will bias the sample selection in some way. But you need to be biased in a very similar way in any case, otherwise (say by using a purely random selection) you'll be left with a selection of samples of which 90% will all be rated 5/5.
My own opinion would be that these extreme cases are valid. I listen to the whole of Kalifornia, so being able to encode the first 5 seconds (the fatboy sample) properly would be important to me!
But, I agree that tuning your codec so that it ace's these rare cases at the expense of being the best at the other 95% of samples would be the wrong way to go.
I'm not really sure what factors make a sample difficult to encode. Fatboy seems pretty clear to me, a sound is made of short pulses which our brain percieves as a continuous tone, so that would be a bastard to encode!! What about other things though? If you knew what a general coder found easy and hard, you could just listen to your own cd collection and pick out a great number samples you would imagine to be hard. If enough people did this, you would end up with the hardest parts of every genre and style. Then, you would practically have generic music, but you have cut out all the easily encoded bits, and just have the more taxing stuff, which could then be used to better distinguish between encoders.
My personal expectation of a good coder is to be able to handle snares and symbols properly. Almost all the styles of music I listen to include them, and mp3 often made them sound 'orrible. I might post some samples of the type of thing I mean, but they'd all be aac's or mp3's as I lost the original wav's.. not sure they'd be of any use for testing except to get my point across
And to ask my original question again, Frank, where did you get those encoders from? Where would I get them from?
[ps. rjam, I disable smileys when I remember to, they still get messed up in quotes tho!]
I wouldn't just favor the samples given a high bitrate by most coders for one simple fact. I like drum'n'bass.
Please, refrain from too much laughter and ridicule :red:
This is come to think of it, a fair complaint against encoders.
Most codecs I've used (lame, aacenc, mppenc) supply a lot less than the average amount of bits to this general type of music. Take for instance:
Roni Size - Brown Paper Bag.
It contains accoustic guitar and high snares (high frequencies are a favourite of roni size), which I'm imagining are among the harder thing to encode and get sounding as nice as the original?
AACenc 2.15 -extreme gives it 158.8kbs
MPPEnc 1.1 --xtreme gives it 144.9kbs
(although LAME gives it justice and gives 194kbs at -extreme)
I would imagine these are all less than average? This has led to files sounding really poor quality because of being given such a low bitrate on lower settings (~normal) I can provide some nasty examples when I have more time. So as I just batch encoded at a certain quality level and then deleted the originals, I later found out shit, I can't trust it to encode at a set quality level. Or at least, my opinion of euqal quality differs from the coders opinion. (and yes, i was stupid to delete the originals before giving the mp3's a proper listening to!)
Originally posted by Mac
(although LAME gives it justice and gives 194kbs at -extreme)
LAME really isn't a good example to use for bitrate comparisons. This is simply due to the fact that it will bloat on anything with a lot of high frequency content.. it's not very efficient at all in encoding this.. so it's not really "doing it justice", it's just simply wasteful due to trying to compensate for a design flaw in the format.
Why should it hurt to tune a codec with extreme cases? I imagine "real music" to be a combination of many of these extreme cases (transient signals, pure tones...) where some are more present than others. So if you have tuned your encoder to handle all the extreme cases well (note that you are tuning all known extreme cases at the same time, without neglecting one in favor of another; I imagine this hard to do), shouldn't it work well on those samples where the extremes are less present (ie normal music, whatever that is)?
I don't know if this holds true, perhaps there is no definite answer. I also think (but can't proove, and probably noone will be able to proove the opposite either) that alot of today's music is very similar sonically. Britney Spears contains many sudden strange noises and synthesizer tones. Destiny's Child use "wicked" percussion (from a drum computer/sampler). And I also believe alot of sound is added in the production process to music that may have once actually been recorded (most of today's rock music).
Heh, so this is today's mainstream music churned out by the producers at 1 hit per minute. Normal music? You decide. I can't really comment on "older" music (I can hear Roberto shouting: Pink Floyd!) since I am not exposed to it as much, but suppose the situation is similar.
Me? I'm a trance head and the places where codecs trip up are just the same as in Britney Spears.
PS: when you say that Frank made an unfair test when he compared castanets encoded with different encoder types and that this sample makes mpc shine and all others suck, well, look at the results again. (Also keep in mind he was using 128k ish bitrates and nowhere did he say he was doing any representative testing)
Originally posted by Mac
Most codecs I've used (lame, aacenc, mppenc) supply a lot less than the average amount of bits to this general type of music. Take for instance:
Roni Size - Brown Paper Bag.
It contains accoustic guitar and high snares (high frequencies are a favourite of roni size), which I'm imagining are among the harder thing to encode and get sounding as nice as the original?
AACenc 2.15 -extreme gives it 158.8kbs
MPPEnc 1.1 --xtreme gives it 144.9kbs
(although LAME gives it justice and gives 194kbs at -extreme)
Damn, that particular sample is one of my favorite tuning samples Now, I can't hear any problems at -extreme (AAC) ?
Also, MP3 needs more bits because of the last scalefactor and limited M/S coding abilities - so it is not good to compare direct bit rate of MP3 and, say, MPC.
Well, I think my hearing is going, because on that song, I can only tell a slight difference between -extreme (158k) and -thumb (65k) ???. But no, I can't hear anything wrong with it at -extreme.
I just wonder why any drum'n'bass tends to be given substantially fewer bits than other music?
I imagine "real music" to be a combination of many of these extreme cases
Hmm, not really. If most music was made up of these hard to code examples, then surely most music would sound pretty awful when encoded at less than 200kbs, like a lot of these samples tend to do!! I only know of two songs that contain anything like the fatboy sample... Kalifornia where the sample comes from, and 18, a song by me (which incidentally, AAC encodes the funny sample bit 30kb higher than the rest of the song)
Originally posted by Gecko
Why should it hurt to tune a codec with extreme cases? I imagine "real music" to be a combination of many of these extreme cases (transient signals, pure tones...) where some are more present than others. So if you have tuned your encoder to handle all the extreme cases well (note that you are tuning all known extreme cases at the same time, without neglecting one in favor of another; I imagine this hard to do), shouldn't it work well on those samples where the extremes are less present (ie normal music, whatever that is)?
I don't know if this holds true, perhaps there is no definite answer. I also think (but can't proove, and probably noone will be able to proove the opposite either) that alot of today's music is very similar sonically. Britney Spears contains many sudden strange noises and synthesizer tones. Destiny's Child use "wicked" percussion (from a drum computer/sampler). And I also believe alot of sound is added in the production process to music that may have once actually been recorded (most of today's rock music).
Heh, so this is today's mainstream music churned out by the producers at 1 hit per minute. Normal music? You decide. I can't really comment on "older" music (I can hear Roberto shouting: Pink Floyd!) since I am not exposed to it as much, but suppose the situation is similar.
Me? I'm a trance head and the places where codecs trip up are just the same as in Britney Spears.
PS: when you say that Frank made an unfair test when he compared castanets encoded with different encoder types and that this sample makes mpc shine and all others suck, well, look at the results again. (Also keep in mind he was using 128k ish bitrates and nowhere did he say he was doing any representative testing)
Castanets and Fatboy aren't very transient. There's much harder to encode (real world music)
out there where these two are dull and adagio-like. Attack times of some hundred microseconds,
50% of the energy above 10 kHz.
It seems to be that the human ear has the best time resolution above 10...12 kHz.
Currently MPC uses 350 and above kbps for such signals.
Frank, I have a question who haunt me for month... I encode my classical music in mpc for month, and I noticed immediatly after leaving mp3 than Musepack had I « strange » behaviour with some instruments. Piano don't need too much bitrate, with mp3, mpc or Vorbis. But a violon (not a critical instrument ) seems overrated by mppenc : +20% (200 on --standard ; 230-240 on extreme, etc...). Harpsichord, organ... the same thing (a bit less fororgan, but harpsichord is more problematic). With --alt-preset standard, I obtain 180 kb/s, and never reached 200 kb/s : mp3 is very cool for classical listener who don't like Metallica. But with Musepack, Mozart need as much bitrate as AC/DC with mp3 encoding :mad:
I recently find a strange and forgotten instrument, called glass harmonica : an horrible and distorded sound !!! Brrr... With --alt-preset standard, an adagio (quiet but awfull music) need only 150 kb/s ! With mpc --standard : 250 kb/s !!!!
???
Why distorded music (harpischord, baroque instruments) are needing so much bitrate, and why heavy metal don't ? Can you, or someone else, help me to understand this big differences ? Thanks a lot
[sorry for my poor expression, and thanks again for your job]
i personaly can tolerate the artifacts of AAC at ~96kbps but not less, i think that hides a little the crappyness of the sound while vorbis accentuates it, well thats just my point of view...
Originally posted by guruboolez
Frank, I have a question who haunt me for month... I encode my classical music in mpc for month, and I noticed immediatly after leaving mp3 than Musepack had I « strange » behaviour with some instruments. Piano don't need too much bitrate, with mp3, mpc or Vorbis. But a violon (not a critical instrument ) seems overrated by mppenc : +20% (200 on --standard ; 230-240 on extreme, etc...). Harpsichord, organ... the same thing (a bit less fororgan, but harpsichord is more problematic). With --alt-preset standard, I obtain 180 kb/s, and never reached 200 kb/s : mp3 is very cool for classical listener who don't like Metallica. But with Musepack, Mozart need as much bitrate as AC/DC with mp3 encoding :mad:
I recently find a strange and forgotten instrument, called glass harmonica : an horrible and distorded sound !!! Brrr... With --alt-preset standard, an adagio (quiet but awfull music) need only 150 kb/s ! With mpc --standard : 250 kb/s !!!!
Why distorded music (harpischord, baroque instruments) are needing so much bitrate, and why heavy metal don't ? Can you, or someone else, help me to understand this big differences ? Thanks a lot
[sorry for my poor expression, and thanks again for your job]
Harpsichord is one of most difficult to encode instruments.
Originally posted by Frank Klemm
Harpsichord is one of most difficult to encode instruments.
One of the nicest, too...
But I don't understand why mp3 find it easy to encode (normal bitrate · 190 kb/s with --aps) and explode with an electric guitar (don't experiment it, but read it many time on many forums). And why mpc choose to increase bitrate on a simple violin (one note during 10 seconds : there is no attacks, but the bitrate jump over 200 kb/s with --quality 5 !), and not on Slayer's cacophony. Is a violin grainy, with a lot of hidden details ? If my supposition is true, thats a good thing if mpc encode it perfectly, and a big progress over mp3 : I want details, even if I can hear them.
But is my supposition true ? I often see a big bitrate jump on distorded instruments : have sub-band encoders problems with theses sounds ? Need mpc or mp2 more bitrate for a good encoding of distorded signals ?
I'm sure you kwow the answers. Thanks for your reply...
As said before, MP3 is not a good format to make quality/bitrate collerations from due to the fact that encoding frequencies over 16khz will cause the bitrate to increase tremendously, even if otherwise this content should be easy to encode. Metal is full of this type of thing.
If you compare a metal encode with aps to a similar quality encode with mpc, aac, or vorbis, none of them will be as high in bitrate... and that doesn't mean they are lower quality either.
As for why MPC increases the bitrate a lot on those other signals, I suspect it is due to it's significantly more advanced psymodel over LAME simply deciding that more accuracy is needed. I'm inclined to believe that this is a Good Thing especially since encoding artifacts are quite rare with MPC compared to other formats.
Again, LAME is really not a good candidate for comparison.
Oh, and theoretically subband encoders are not as efficient in encoding some signals as transform coders (frequency based vs time based), but I don't believe that is actually what is causing the bitrate increase here, because certainly the jump would not be that high... not in the 200kbps+ range at any rate. Of course, I could be wrong
Thank you very much, Dibrom, for your answer.
LAME is maybe not a good candidate for comparison, but I'm a bit perplex when I see the different behaviour between audio codecs. As I said it on an another topic, critical sample are making codecs crazy (+100 % for mpc standard, -66 % on Vorbis, and CBR who are not CBR at all...). In that case, this is funny. But in real musical tracks (I apologise for the people who listen every morning short_block.wav), it is interresting. Isn't it more clever not to choose only ONE codec for encoding CDs, but choose the better one for each instruments ? MPC for electronical, Vorbis or AAC for violin, and mp3 for glass harmonica ? Or even change for each movement (I noticed that quiet movements : largo-andante... needs much bitrate than vivace-allegro-presto...) ? Not very practice. But strange for a guy who was familiar with mp3 behaviour.
A good psymodel should be able to cope with almost any situation thrown at it. I don't think a different encoder is needed for every type of music.. that approach makes no sense IMO. Futhermore, unless you've actually heard a problem with a codec like MPC for example (which you can provide samples and abx scores for), then why bother worrying about the bitrate it chooses on a particular sample? (given that the average bitrate overall is acceptable, which I believe it is)
Let the psymodel decide what it thinks is right, and let your hearing do the rest. Trying to analyze quality or interpret results based on bitrate and situation is pretty much just as meaningless as looking at spectrograms and trying to decide which codec is best that way.
i will try to answer a few questions here... mind you im not an expert on the subject.
first off, slower more quiet movements need more bitrate, because there is less volume in general, and thus less volume of noise and sound to cover up other sounds. as such, the encoder keeps more information about the sound, because more of it must be played back, as less of it is "covered up" by louder noise and sounds. (i hope you can understand this... and i hope i am correct at least partially on this)
as for mpc having higher bitrates, i believe you are correct in the idea that harpsichord, violin, and glass harmonica, while seemingly simple, are actually quite complex. i personally have some tracks by blues traveller, in which john popper plays a normal harmonica, and when it is solo and lower volume, the mpc at standard --ltq fil jumps to over 230 kbps. later when it is "covered up" by drums and guitars and some audience noise, the bitrate drops down to about 190 kbps or so. i would believe this to be another example.
one last thing. this same sort of "phenomenon" can be seen on the encoding of live music. when the audience applauds, the encoder of course does not keep the sound of every single hand clapping, but (in the case of a large audience) there are so many hands clapping that there is a tough job to be done when deciding what to keep and what to get rid of. i have noticed that mp3 does a very bad job, and the applause sounds like the ocean (very swishy) while mpc handles it fairly well (albeit alloting a lot of bitrate to a seemingly simple sample). this is perhaps less obvious, as when you hear applause in a track, you tend to ignore it, or your own mental "image" of what applause sounds like takes its place, and you associate that sound with the sound coming from your speakers.
i hope i am correct in what i am saying, and i hope i have helped you with your answers, guru
Originally posted by Dibrom
Futhermore, unless you've actually heard a problem with a codec like MPC for example (which you can provide samples and abx scores for), then why bother worrying about the bitrate it chooses on a particular sample? (given that the average bitrate overall is acceptable, which I believe it is)
I totally agree with you. Harpsichord CDs are compensated by piano CDs. For me its not a problem, and I don't really care about. But as I often read it on this forum, space ALWAYS matters, no ?
Originally posted by spase
first off, slower more quiet movements need more bitrate, because there is less volume in general, and thus less volume of noise and sound to cover up other sounds. as such, the encoder keeps more information about the sound, because more of it must be played back, as less of it is "covered up" by louder noise and sounds. (i hope you can understand this... and i hope i am correct at least partially on this)
I understand perfectly. It make sense... it's just that mp3 seems to have the opposite behavior, and I was « educated » with another model.
as for mpc having higher bitrates, i believe you are correct in the idea that harpsichord, violin, and glass harmonica, while seemingly simple, are actually quite complex.
Yes, of course, I can feel this complexity. And hear bad distorsion on harpsichord encoded with mp3 (--r3mix, and --aps) : bitrate is low, quality too...
i personally have some tracks by blues traveller, in which john popper plays a normal harmonica, and when it is solo and lower volume, the mpc at standard --ltq fil jumps to over 230 kbps. later when it is "covered up" by drums and guitars and some audience noise, the bitrate drops down to about 190 kbps or so. i would believe this to be another example.
Same thing for a violin concerto (or an orchestral work with a solo violin at the middle of the track) : +20-30% for a solo instrument, less for the 120 others playing at the same time.
i hope i am correct in what i am saying, and i hope i have helped you with your answers, guru
Very well, spase. Thanks a lot for this answer.
But don't blame me if I request again the opinion of the mpc developer : just for curiosity.
Thanks a lot for all answers
Originally posted by guruboolez
But as I often read it on this forum, space ALWAYS matters, no ?
lol
spase always matters? :diabolic:
Same thing for a violin concerto (or an orchestral work with a solo violin at the middle of the track) : +20-30% for a solo instrument, less for the 120 others playing at the same time.
anyhow, yes as in the beginning of the first movement of rimsky-korsakov's sheherezade ("the sea and sinbad's ship" in english)
or perhaps part way through the russian easter overture, by the same composer (one of my favorite composers indeed)
anyhow, i'm glad to help...
by reading your other posts, its obvious your ears surpass mine... i have an interesting sample... i used it once when i was trying to decide between mp3, vqf, wma, and the three new (at the time) formats of ogg, aac, and mpc... at the time i weighed in factors of encoding time, decoding speed/cpu usage, average size, and of course audio quality, and i found mpc to be the most favorable when concidering all options.
if you are interested i could send the file to you via email to see if there are any artifacts left in it... im sure it would be best to test on quite low bitrates... i know this is getting a wee bit off topic but hey whatever...
email or send me a private message here on the forum if you are more interested in the file...
edit: lol by now youd think i would know how to format these posts!
Very well, spase. Thanks a lot for this answer.
But don't blame me if I request again the opinion of the mpc developer : just for curiosity.
he doesnt visit this forum THAT much... better to drop a email his way...
(sorry for double post)
edit: once again the formattting....
Originally posted by spase
first off, slower more quiet movements need more bitrate, because there is less volume in general, and thus less volume of noise and sound to cover up other sounds.
So... Can we guess these are complicated situations for time[/b] domain based codecs? (like MPC)
Because of the format's structure, it's well suited to encode and confine transients/space/bitrate but ... for long evolution signals it has problems to
shortly synthesize the real "body" of sound: the frequencies?
Any voluntuer can briefly explain (I don't require simply words, without commitment, guys) to what consist the
time based coding of MPC,
Time... in which sense? Shouldn't it encode freqs too, as the others DCT based codecs?
again im not an expert, but assuming the temporal compression (i guess thats what you are talking about) is similar to that used in movie compression, i would guess when adjacent frames are not even close to being similar it would provide a problem.
i dont know all that much about the insides of the codecs... but i try
Originally posted by unplugged
So... Can we guess these are complicated situations for time domain based codecs? (like MPC)
IIRC the only time domain based codecs are the lossless ones. Since Musepack performs spectral analysis and subband encoding, a Fourier transform is somewhat of a necessity. You can't speak of i.e. 0-200 Hz frequency band when in time-domain, can you? Time domain usually involves things like Huffman codes, RLLs, convolution and such.
Now that I'm at it, why does everyone keep saying that codecs using DCT are the only "transform" codecs whereas codecs like MPC (using FFT) are "subband" codecs with nothing to do with transforms at all?
"Subband" encoding does use discrete Fourier transform. "Transform" encoding uses discrete cosine transform. Mathematically speaking, these transforms are nearly identical, with cosine transform being nothing but a cosine-termed (is this the correct english expression?) Fourier series (Fourier transform without the cosine, or ImX, part). Cosine transform just makes energy representation a little easier than Fourier transform. The differences of subband and transform codecs are
much more profound than a simple variation in a transform equation.
Well, so much for pathos
About the glass organ which was encoded at only 150kbps by mpc. The recent versions of mpc have included code that takes better advantage of stereo correlation between channels and thus can compress tracks with minor stereo separation alot better. (Allthough they don't sound "mono-ish" to me). I can't tell beforehand by listening to a track, if the stereo correlation can be exploited well or not. I'm often surprised by what the encoder does but I can't hear any flaws. There has been a long discussion about this on the phorum (http://www.chaostar.org/phorum/read.php?f=1&i=3892&t=3892), but beware: the method how Kevin T. analyzes codecs is flawed! Do not be mislead by his posts!
Now I'm not sure if this applies to your glass organ sample, but it could be a possible explanation why such a tonal instrument gets encoded at such a low bitrate. But hey: if it sounds good, then don't worry about the bitrate.
I stop you : mp3 gives 150 kb/s and mpc 250 kb/s
And I discover the problem before mppenc 1.00 (the 0.90 era). Frank Klemm has nothing to do with it
Last thing : it's glass harmonica, not organ,a very strange and rare instrument :
Possibly this is all nonsense but here goes anyway:
A subband codec only splits up the signal into, well, subbands. Because each freqeuency band can be encoded at a different resolution/sampling rate (see Nyquist). This allready reduces the amount of bits needed to represent the signal. This process is not lossless but nearly. I don't know how this is done but I see no way around Fourier transformation or something similar. Then the signals of the subbands are quantisized. Here you are operating in the time domain again. The quantisation process is where all the psymodell magic happens.
A transform codec first moves everything into the frequency domain usually by mdct. Then the mdct coefficients are quantisized (not the signal itself). So here you are working in the frequency domain. This works somewhat more efficient so you can generally achieve lower bitrates while maintaining the same quality. The mdct is where preecho can happen, because when converting to the frequency domain you have to work on a rel. large number of samples (or better: over a period of time, which can not be infinitessimally small).
When you do the conversion from the time domain into the frequency domain you allways have to make a compromise between time resolution vs. frequency resolution. There's a nice physical example which illustrates this:
You can build a device to measure the frequency of a sound by having several metal prongs of different lengths alined next to another like a comb. each of these prongs is resonant at a different frequency. Now imagine you play a sound. Sound waves are basically swinging impulses of different air pressure levels. The speaker which is playing the sound moves the air at a certain frequency. But the very first impulse of air hits all prongs the same way. They will all start moving at their own resonant frequency. When the second impulse comes along some prongs will be affected more than others because their own movement correlates with the movement of the air. This goes on and on and in the end in theory only one prong will be left swinging which matches the frequency of the speaker.
The point is: it takes some time until you can determine what frequency is playing by looking at the prongs. If you were to decrease the number of prongs this would make the process faster because the resonant frequencies of the individual prongs are further apart. So you have increased the time resolution but had to sacrifice frequency precision.
If you increase the number of prongs (higher frequency resolution) it will take longer to deteremine which frequency is playing because some prongs whose resonant frequency is close to the one being played will take longer to stop swinging. Thus you loose time resolution.
This is not all bad because our ear works in a similar way. I hope you can understand all this as I would be better explaining this in German.
I stop you : mp3 gives 150 kb/s and mpc 250 kb/s
And I discover the problem before mppenc 1.00 (the 0.90 era). Frank Klemm has nothing to do with it
My apologies for my mistake! Mpc's high bitrate is fairly easy to explain then. When you quantisize the signal and loose some of the info you introduce noise. This noise has to be hidden somewhere among the other sounds so you can't hear it. If your sample is very tonal there is not much space to hide the noise because the human ear hears the difference easily. Ergo you have to quantisize with higher resolution to introduce less noise and this in turn raises the bitrate.
Mp3 on the other hand only has to store a few of the mdct coefficients, because most likely the signal will be made up of just some simple waveforms + overtones. (See my other post above)
Just for the record: 0.90 is allready the Klemm era, but in the beginning most optimizations were related to speed.
Last thing : it's glass harmonica, not organ,a very strange and rare instrument
Well, in German it's called "Wasserorgel" where Orgel = organ, just me mixing things up.
Cheers
Thanks guys,
Gecko very well explained the matter
So, a NOT indifferent lack point of transform codecs is the response time!! What a lack!
It's not a great news to known today...
mmm... for example, with transform codecs we cannot record a certain freqency that raises or decrease (the slide), we can "only" record a X frequency at Y time respecting the Z time resolution/granularity (compromise).
Thanks again for the interest, must say this mostly happens only at HA.
Now that you're talking about subband vs. transform, I am wondering why nobody is taking the advantages of each method. How about an hybrid codec that would decide to encode each frame either as subband or transform?
Maybe frames with strong transients could be encoded as well as mpc does, and more tonally simple ones could be encoded efficiently as with aac or vorbis...
Of course it would require twice the effort in coding and tuning and it would encode slower. Do you think that bitrate reduction would justify the drawbacks? Maybe vorbis could implement subband in 2.0 ?
Maybe this is all nonsense?
Jaime.
Originally posted by jalonsom
How about an hybrid codec that would decide to encode each frame either as subband or transform?
That codec is called MPEG Audio Layer 3.
Ever heard about MP3's hybrid filterbank?
Originally posted by rjamorim
That codec is called MPEG Audio Layer 3.
No, MP3 _always_ does _both_.
Switching between the two is problematic because transform codecs use overlapped windows, i.e. you can't really encode frames independently.
--
GCP
blah, AAC beats MP3 it being a transform codec
Maybe vorbis could implement subband in 2.0
Afaik "Subband Technology" is unfortunately patented by Philips. See here (http://www.audiompeg.com/patents.htm), number 8.
Originally posted by Gecko
Afaik "Subband Technology" is unfortunately patented by Philips. See here (http://www.audiompeg.com/patents.htm), number 8.
Subband coding is NOT patented.
Read patents very very carefully or don't read it.
Perceptional noise substitution is also NOT patented.
When reading patents it is necessary to find out what is EXACTLY patented.
Maybe vorbis could implement subband in 2.0
That was not the main bone of contention in the first place, Monty had decided NOT to use subbands for reasons based upon computational complexity and any other reasons he may have came to conclusions with. It also paved the way for more experimentation that could be performed, just as maybe theoritcally speaking Hybrid Discrete Wavelet Filterbanks which I try to outline so fondly as many ask.
:gah:
Originally posted by Frank Klemm
Subband coding is NOT patented.
Read patents very very carefully or don't read it.
Perceptional noise substitution is also NOT patented.
When reading patents it is necessary to find out what is EXACTLY patented.
Sorry, I am not able to understand documents where a large amount of the sentences spans more than 40 lines of text. Those of you who can, can read the patent here http://www.depatisnet.de (http://www.depatisnet.de). English pages are available. Search for Patent "EP 0400755 B1". After reading parts of the description and of the claims, I don't see why musepack does not fall under this patent. Please don't tell me to read the whole thing (and understand it ). Are you aiming at the transmitter/reciever part? Imho this is covered in the claims.
Usually a court will decide what exactly is patented or not .
What is the difference between the mechanisms described in the patent and the ones implemented in mpc?
Where does mpc use patents then?
Subband coding is NOT patented
---
Philips subband patent can be removed
Now I'm confused. ??? I'm sorry if I didn't use the proper, exact terminology.
Way back in third grade a guy once said: "I didn't spit at her, and besides, I missed!"
Originally posted by Gecko
Now I'm confused. ??? I'm sorry if I didn't use the proper, exact terminology.
Way back in third grade a guy once said: "I didn't spit at her, and besides, I missed!"
I think what he's saying is that subband coding in itself is not patented, but the specific way it's done in Musepack now is.
--
GCP
Back to my original point... currently AAC is pretty bad at low bitrates, and MPC is worse than AAC, I'm not sure how MPC could stick around as a major format because it can't get that much better at low bitrates, which is how alot of stuff is transferred on the internet. (eg, I encode at -internet (~70k) for giving my music to friends, would like to use -thumb (~50k), but that just sounds horrible!)
Do you reckon -thumb would start sounding reasonable with the AAC+SBR? I find -internet is high enough quality to listen to without being continually reminded that its nasty quality.
I think the short-block mode of AAC is giving the problem. Long Block is more efficient than short block. In fact, for most complex musical clips, the PE measurement is just about 600 ~ 800 bits but when there is a lot of signal transients, the AAC algorithm has to switch to short block, which requires 2~3 times more bits than long block. Worse, the length of the long block mode, 2048 time samples would mean that when switch to short-block, more time audio sample would have to be coded in short-block. In fact, the AAC short block is more inefficient than the MP3 short block
mode. AAC requires 8 short block, 256 time samples each whereas MP3 only requires 3 short block, 384 time samples each. Also, MP3 specs allowed block switching on a frequency basis, eg: frequency above 2Khz coded in short-block while anything below can be coded in long-block. This is a feature not available to AAC, not even for the Gain-Control tools.
For long block, there is a theory that states that maximum block length for most efficient coding is about 2048 time samples. Anything above or below this length would require more bits. There is alot of active research in abolishing the need to switch to short block such as switching to wavelet filter banks or the Gain Control tools during signal transients.
Also, there is alot of research into noise-tone classification model which provides even more coding gain which I believed the MP3Pro is based on. However, how good the audio quality at Hi-Fi level is unclear. I tested MP3Pro on the castanets clip and it seemed to contain some irritating artifacts. I would not classify MP3Pro as a Hi-Fi / CD level encoder.
wkw
I think the short-block mode of AAC is giving the problem. Long Block is more efficient than short block. In fact, for most complex musical clips, the PE measurement is just about 600 ~ 800 bits but when there is a lot of signal transients, the AAC algorithm has to switch to short block, which requires 2~3 times more bits than long block. Worse, the length of the long block mode, 2048 time samples would mean that when switch to short-block, more time audio sample would have to be coded in short-block. In fact, the AAC short block is more inefficient than the MP3 short block
mode. AAC requires 8 short block, 256 time samples each whereas MP3 only requires 3 short block, 384 time samples each. Also, MP3 specs allowed block switching on a frequency basis, eg: frequency above 2Khz coded in
Originally posted by wkw
I tested MP3Pro on the castanets clip and it seemed to contain some irritating artifacts. I would not classify MP3Pro as a Hi-Fi / CD level encoder.
mp3pro bitrate limitation is very low. At this bitrate, no codec can claim CD-quality. At least for a lot of samples, and for most music.
But a agree with you : mp3pro spec, based on mp3 spec, is limiting the theorical quality on transient signal. It is not really CD quality, even on -b 640 --freeformat.
EDIT : I don't understand anything on lossy/lossless specs. I just experiment these preecho problem on mp3/mp3pro samples.
Bah, I hate it when this happens. I was told off for getting cross-subject in one of my threads, so keep ya MPC whatnots to yoursleves!
If the shortblock is inefficient, could it be improved drastically? Or maybe it's like that for a reason (eg, giving better quality than mp3 shortblocks ever did)
Originally posted by niktheblak
Now that I'm at it, why does everyone keep saying that codecs using DCT are the only "transform" codecs whereas codecs like MPC (using FFT) are "subband" codecs with nothing to do with transforms at all?
"Subband" encoding does use discrete Fourier transform. "Transform" encoding uses discrete cosine transform. Mathematically speaking, these transforms are nearly identical, with cosine transform being nothing but a cosine-termed (is this the correct english expression?) Fourier series (Fourier transform without the cosine, or ImX, part). Cosine transform just makes energy representation a little easier than Fourier transform.
Because the FFT in subband coders like MPC is done in psychoacoustics calculations which is a separate process. Psychoacoustics defines the masking threshold which is used in the quantization phase. With subband codecs transform coefficients are not used in actual encoding/quantization. Transform codec like MP3 also uses FFT (normally) for psychoacoustics, but as I said it's a separate process...
MPC quantizes the time-domain samples (based on the masking threshold given by psychoacousic analysis), not any frequency transform co-efficients like transform coders.
Thread splitted.
Some MPC specific messages have been moved to the MPC general forum:
http://www.hydrogenaudio.org/forums/showth...=&threadid=3068 (http://www.hydrogenaudio.org/forums/showthread.php?s=&threadid=3068)
wkw,
AAC has very efficient mode of grouping of 8 short blocks into groups which in most cases reduces necessary bits by a significant margin.
Basic AAC does not have SBR tools implemented in, so for compare with mp3pro we would have to wait AAC+ (MPEG-4 V3) and see how does it match with mp3pro. CT (codingtechologies) already stated that AAC+ is superior to mp3pro.
Originally posted by JohnV
Because the FFT in subband coders like MPC is done in psychoacoustics calculations which is a separate process. Psychoacoustics defines the masking threshold which is used in the quantization phase. With subband codecs transform coefficients are not used in actual encoding/quantization. Transform codec like MP3 also uses FFT (normally) for psychoacoustics, but as I said it's a separate process...
MPC quantizes the time-domain samples (based on the masking threshold given by psychoacousic analysis), not any frequency transform co-efficients like transform coders.
Subband and transform encoder do a time decimation and split the signal into multiple
bands. I would call a subband filter a "multiple overlapped transform".
Subband encoders: multiple overlapped transform
- filter length: 512 (first zero cross at +/-55.9)
- decimation: 32
- overlap count: 16
Transform encoders: dual overlapped transform
Long block AAC:
- filter length: 2048 (first zero cross at +/-1023)
- decimation: 1024
- overlap count: 2
Short block AAC:
- filter length: 256 (first zero cross at +/-127)
- decimation: 128
- overlap count: 2
Originally posted by Ivan Dimkovic
wkw,
AAC has very efficient mode of grouping of 8 short blocks into groups which in most cases reduces necessary bits by a significant margin.
Well, from my observation on the sum PE measurement of 8 short-blocks with PE of 1 long-blocks, it seemed that even window grouping of short blocks will not reduced the bits required to code in short blocks to the level of long block mode. Window grouping in my opinion only reduces the "side info" of short-block.
wkw