I already did that, but I wasn't sure if any corrections were needed. Here's what Chunky calculated (please correct me if the values are not right):



and here's the chart:

This looks much better now. Nero is at its place
The only way to check values is to compare them to those that Mares published.

That is actually not a contradiction as such (though further expert opinion on the actual statistical metric used is needed).

You think that is a contradiction, because such situation doesn't happen in "normal" number systems, like integers, reals, etc. What you noticing is the property of total order breaking. However not all valuations have that property.

Take integers, if you take 2 integers at random there is the way to count from one to the other, precisely because there is a total order and you know what is less/greater than what, what equals what and what follows what.
On the other hand take Complex numbers, this is the first number system students usually exposed to in school that doesn't have total order on its elements (though school teachers don't usually mention that). Given 2 random Complex numbers there isn't "the" way to count from one to the other, in fact there are infinitely many ways, all correct in some sense.

So while I don't know if the underlying statistical measure produces set of values that has total order, your example (if not subject to some freaky error) shows that it doesn't, and should be read as raking:

1) HE = WMA
2) Ogg

My immediate intuition would be to use equivalence classes to solve this problem.
1) Make individual comparison between every possible pair
2) Look at the ones with strict inequalities
3) Pick the largest of them all (strictly greater, not >=)
4) Rank that first
5) Add all who are directly equal to it to its equivalence class (not ones that are equal by some chain of equalities)
6) Removing them from further consideration
7) From remaining, rank the next largest as number 2
8) Go to 5 and repeat for the rest of the ranking.

How about use approximation instead of equality.

Same opinion for me.
For clearly defined objects A, B, C, a clearly defined identy and a clearly defined <-relation it would be a contradiction.
Here A and B correspond to the quality of Nero HE AAC and that of WMA pro, and C corresponds to that of Vorbis. Quality as measured with this test.
The problem is in the meaning of '=' and '<' as these are rough quality comparison operators which can easily make up for such a pseudo-contradiction.

The zoomed view is a major evil to me as it overestimates such a rough '<' comparison.
From absolute view ranging from 1.0 to 5.0 it's easy to say 'all these three encoders yield roughly the same quality with vorbis being a tiny bit behind.'
This is what is most important in practice, cause with these codecs you usually don't have the choice which one to use on a mobile device. No matter which one you use: you get state of the art 64 kbps technology regarding quality.

If it's up to elaborating differences between the encoders the very personal preferences are much more of concern than the overall small quality differences according to the test.

You can not assume transitivity in test results.

Interesting results. I only use Vorbis for high bitrates (my own music collection on my PC's hard disk), since I expected low bitrates (for a small, portable mp3 player) to kill Vorbis. Seems like I was wrong.

With those results, low bitrate Vorbis internet streams make more sense now!

Wouldn't it make more sense just to go with one of the winners for low bitrates? (Also note that on some of the subtests, ogg did much more poorly relative to the others.) Or are there other tradeoffs here?

Well, for streaming purposes you're right, 64 kbit/s WMA10 Pro/Nero AAC seem to be better choices than Vorbis here. For low bitrates in general this isn't implicitly the case, because Vorbis, unlike the other two codecs, doesn't rely on tricks to artificially improve the quality. HE-AAC includes Spectral Band Replication, WMA 10 Pro makes use of a very similar approach. For portable players these techniques can be quite a burden, since they heavily drain batteries. Of course, Vorbis isn't a saint concerning its power hunger as well, but judging from what I've read so far it isn't as demanding as HE-AAC. I haven't seen any figures about WMA 10 Pro's decoding performance on low bitrates yet, hence I'll abstain from further commenting it. A Zune owner might be able to shed some light on this matter.

Thanks, I forgot about that whole dimension to the problem. It's like comparing compression utilities by size reduction and forgetting about speed. It would be useful to have some power consumption index, but I guess that may vary greatly by device.

Out of curiousity, why is there no castinettes in this test? In my recollection pre-echo was a large problem with several of the codecs, and not with some others. It would seem unreasonable to suppress this issue.

AFAIR, it's "castanets", not castinettes, and there are other samples in it that show that issue. You'll definitely find comments about preecho in several test comments (i know, i wrote some)

Edit: typos.

Indeed there are some other sources, but these other sources also have ring tones and such, which mask a lot of pre-echo. I don't doubt you could still hear some. Oh, and thank you for offering the spelling corrections, I am elucidated. Yeah.

Well, I asked for sample suggestions long before the test started...

How many of you that took test did use Headphones? How many did use in ear phones?

headphones for me

Cheap Sennheiser HD 60 TV Headphones an AC97 onboard Soundcard.

I was recently thinking of an additional questionnaire in future listening-tests, e.g. for listening environment, age, ...

Headphones for 17 samples. Loudspeakers for White America samples because I couldn't hear the diff in headphones. Masking, noise canceling?

good headphones (mid-price Beyerdynamic)

with some generic on-board sound on a laptop.

Headphones: Koss PortaPro

I too used the Koss PortaPro headphones, which are quite suitable for testing encoders at this quality level.

...and as I wrote before:

These speakers have a price tag of about EUR 2000. The PortaPros are about EUR 50 or less.





My comments about White America:

"I hate this sample. My previous sample was the quiet bibilolo and I had set the volume level louder than normal. I didn't remember to reduce the level before starting this. It was like an explosion inside my eardrums. I hope I didn't damage my hearing...

In general the sample is overcompressed and very distorded. It does not have much that encoders could hide or alter. A bit more pre-echo & distortion does not change the ugly nature of this sample."

It seems to me that in most cases headphones reveal more differences than speakers and that most of participant in public listening tests use headphones.
Also while doing this listening test, I discovered that in some cases I could hear more differences with cheap earphones Creative EP 630 than with Sennheiser HD 650. I guess it is because of blocking of outside noise (though there was not much outside noise since I was in a room with doors and windows closed, and using quiet HTPC).
Or maybe it has to do something with neutralizing effects of head and pinnae related filtering.

Could it be because the cheaper ones don't produce all frequencies as well/evenly, which reveals high frequency artifacts better...?

It is true that these headphones have different frequency responses:
http://www.headphone.com/technical/product...ild-a-graph.php
http://www.hydrogenaudio.org/forums/index....ost&id=3535
But it is question if this is the only thing that causes different perception of differences.

And difference between discovering artifacts with speakers or headphones IMO is beyond just different frequency responses of speakers/headphones. There are examples of double talk artifacts that sound terrible on headphones and are hard to perceive on speakers. One for example is sample 16 in this test.

I used a pair of Sennheiser PX 100s for the test.



I composed something similar but deleted it prior to submission due to unprofessionalism. Of all samples I spent significant time on I had the most difficultly with that one, precisely because the sample itself was so offensive to my ears. Even with volume levels down the song remains a screeching blob.

All of a sudden, I have got a small question -- about the error bars on all the plots.

If we compare the plots for two different samples, the error bars are shorter for the sample with more listeners. This makes sense. (More listeners --> more representative statistics --> less error) Ok.

But if we look at just one plot (any one of the plots), it seems the error bars of all 5 contenders have exactly the same size. Are they actually exactly the same? Is it how it's supposed to be due to the design of the test?
Are there any circumstances when error bars could have different size for different contenders?

Within a sample plot, all bars should have the same size - always.

Somehow this feels counter-intuitive.

Imagine an extreme case when one contender is rated 3.0 by ALL listeners (i.e. all of them give exactly the same rating), but other contender gets different ratings between 1.0 and 5.0
Why should the error bars be equal?

(I don't doubt the results, just want to understand a little deeper.)

Maybe someone with more knowledge in statistics can answer your question.

Who said all bars should be equal? What do you want the bars to represent?

some boxplot example: http://www.physics.csbsju.edu/stats/box2.html

In my results (and Roberto's, Guru's and ff123's), the bars for the various contenders of the same sample will have the same length.

If the bars are supposed to indicate the quartiles they should vary a bit. But I haven't checked what those bars are supposed to be...

For this type of analysis, the error bars are all the same size. Another way you can do the analysis is to have a different confidence range for every comparison. So for the 5 codecs (including the anchors), you would have 10 different numbers. This can be represented well in matrix table format, but not nicely in a graph format. If you want to get matrix type confidence ranges, download the bootstrap program from my site:

http://ff123.net/bootstrap/

which performs this type of analysis. In practice, the two types of analyses yield very similar results.

So the bars do not represent the distribution of data collected for each codec, as, for example, you could have one codec rated by all people 5.0 and you'll add bars to it. I find this confusing. What is the meaning of the painted bars? How should I read them?

I would've loved to see MP3 involved in this test. We know that Vorbis, AAC, and WMA are better, but just as a comparison it's always interesting to see how the newer, improved codecs rate nowadays against our friendly ol' MP3 fomat, to know exactly how much of an improvement there is.

If the bottom of the bar of one codec does not touch the top of the bar of another codec, you can state with at least 95% confidence that the first codec is better than the second one.

The bars being all the same size means that you might lose a bit of power in making statistical distinctions between codecs. But I think that's more than balanced by having the nice, easy-to-look at pictures instead of tables of numbers.

There are some who assert (and they have a point) that even if there are statistical differences between codecs, it may not make a practical difference if the ratings are relatively close to each other (close being determined by looking at the pictures and making a judgment).

Sorry for bringing this up again, but I have one more note about this. iTunes 96kbps VBR was used in this test at 64kbps and in previous at 48kbps. Some samples were used in both tests. But score for those sample is not the same (example: Toms Diner 4.70 vs 4.86) and the decoded sample is the same. Even a participant involved in both tests didn't give the same rating (examples: Alex B 4.0 vs 4.2, haregoo 5.0 vs 4.5).
Unfortunately it is not possible to get consistent results

Yes, this is normal and depends on the mood, the listening-conditions (maybe different headphones or soundcard, possible noise from the neighbors, etc.) and health (maybe the listener just got better from a cold or still has a cold while testing).

Say, are there any plans for doing a new test here?

As I mentioned elsewhere, it's not a good apples-to-apples to compare CBR WMA to qulaity VBR in other codecs. The WMA family supports quality VBR, as well as 2-pass CBR and bitrate VBR modes.

And for a streaming test, CBR is really the appropriate encoding mode. While fixed quality is an interesting thing to look at, it excludes rate control, which is a very important part of codec design, and a place where a lot of engineering effort goes.

I would agree here. Streaming is the main use so far for 64kbps. Low bitrates are interesting for portable devices, but the CPU usage (and hence battery life) of the winners of this test (HE-AAC and WMA Pro) leaves a lot to be desired.

How are you measuring CPU use/battery drain of the codecs? We've done a ton of work for the mobile implementations of WMA Pro to get the CPU hit low enough to make it feasible for phone use. I haven't done any formal testing with recent devices though.

The reason why WMA was tested in CBR mode is that Microsoft seems to recommend CBR over VBR for WMA. Also, IIRC, VBR produced target bitrates that deviated from the average bitrate of the other encoders by more than 10%. 2-pass modes for short samples are also not an option - using 2-pass must be done on complete tracks and then samples have to be extracted out of the encoded full tracks.

A pure CBR test could be interesting for streaming indeed.

Do we? Do you have a link - I'd like to have that corrected. Speaking for Microsoft, I recommend that content that needs CBR be encoded as 2-pass CBR, and otherwise 2-pass VBR be used. We've done a lot of work around 2-pass audio encoding.


Hmmm. How short are the clips you're using? If you can give me a reproducible test for this, I'll pass it on to our engineers. In my experience, VBR audio comes out within 1% of the target, but I'm normally encoding at least 60 second clips.

2-pass VBR peak limited might work better in this case. But if you need to use CBR, at least use 2-pass.


Great, I'd love to see that as well.

For the WMA codecs, the proper mode to use for that (unless it's a test of live encoders) would be 2-pass CBR. We are able to get a meaningful reduction in peak QP with 2-pass CBR.

The test performed by NSTL featured WMA in CBR mode. Since you explicitly instructed NSTL what settings to use, one would assume you had a reason why you did this: obtain best quality results.

If that is not the case, well, sorry. IIRC, WMA did not offer a quality based VBR mode that produced files with the target bitrate.

Could you explain me what multi-pass CBR is supposed to do? I thought multi-pass encoding was good for ABR only. For CBR you always assign the same number of bits (don't know if WMA has something like a bit reservoir -in case it does, I imagine that could be the only variable thing that could be influenced by multi-pass encoding).
As for bitrate based VBR (which I call ABR) I would prefer to encode full tracks and then extract the sample from that. Otherwise the test has no or less usage.

That test was done before my time, but my understanding is that we used 1-pass CBR in that case as that was the only rate-controlled mode supported by HE AAC, and the goal was to have an apples-to-apples test. It was never meant to be a demonstration of best practices. 1-pass CBR is certainly the most challenging codec mode, so it's interesting to test, but nothing I use other than for live encoding.

Understood. I just want to help make future tests a more scenario-relevant comparison.

Correct. with 2-pass CBR, you're able to essentially request a bigger bit reservoir in advance of complex audio, to keep worst-case QP lower. With 2-pass VBR, we essentially calculate the QP that will produce closest to the optimum bitrate, and then vary QP's per block a little in order to hit the target. But in essence an unconstrained 2-pass VBR is a lot like a "magic" way to figure out what quality level to use to give a file of the requested size.

Makes sense to me.

Moderation: Fixed quotes.

Hi Ben,

Nice to see you here at HA. I think you'll find this place somewhat subdued compared to AVSF..

Interesting you speak of 2-pass VBR WMA. I have been using -a_codec WMA9STD -a_mode 3 -a_setting 128_44_2 for more than a year with excellent results on my portable. I think perhaps it is underrated/underused in the lossless community, though it wasn't trivial to get the VBS command line options all sorted out. The reason I ended up with ~128kb 2-pass VBR WMA was that during my testing, I found it maintained the best stereo imaging during intricate percussion/cymbal passages.

I tried 1 and 2 pass CBR wma10 at 64 kbit/s in past. I didn't share the results here. There were miscellaneous changes but I couldn't abxed the difference.
So maybe 2 pass has a bigger reservoir and other kind of grass called "magic" it makes no sense for audio CBR encoding. If anyone doesn't agree provide samples where 2 pass CBR is better than 1 pass for wma10.

Thank goodness !


Cool, glad it's working out for you.

I'd probably recommend using -a_mode 4 and set a peak bitrate instad of leaving it entirely unconstrained, since devices may have a maximum supported rate. For Zune, it's 320 for audio-only files, and 192 for soundtracks in WMV files, IIRC.

Stuff like stereo seperation is a great thing to use VBR for, since it gets you the bits were you need them. I think people spend so much time sweating the hard clips they can miss that most of most full tracks aren't that hard.

I'm still anxiously awaiting the forthcoming ~80kbps multiformat test, especially now that Ayoume has just released beta 5.5 to infuse more life into Vorbis.