I tested some times ago the lossy compression of wavepack. I found it really nice, and not to easy to ABX.
Today, I tried with one sample, very sharp (percussion symphony of Rolf Lieberman, called Les Échanges), and on the first attempt, I noticed a brighter sound. ABXed easily. Luck ? Am I in a good mood ? Can someone try it ?

http://membres.lycos.fr/guruboolez/AUDIO/s...rman_intro.flac

Tested with -b250 -h setting (=> 264 kbps file)

Compared with mppenc 1.15r --standard --xlevel

Wavepack = 252 kbps
Musepack = 157 kbps



Musepack is easy to ABX : distorsions on attacks (it reminds me the piou-piou I heard with castanets.wav)

Thanks for the sample guruboolez.

I'll check it out in a day or two. I suspect I will be able to ABX it regardless because I always seem to be able to with Wavpack 265/320. (Background hiss)

What are you hearing that lets you pick the difference? Is it background noise, or some artifacts?

I see it trips up Musepack at q5 as well...

Den.

Hiss. But I'm surprise to hear it so easily. I didn't remember, in my last test, to succed with so ease in ABX test.

For musepack, mppenc.15r (tested here) seems to be worse than 1.14. See here.

I've just tried to ABX the 400h mode, and opposed it to mpc --quality 10 :



Other, just after this one :



Is 12/16 really significative ?
On two tests, I obtained :
· 18/32 for musepack -q10 (291 kbps)
· 24/32 for wavepack -b400 -h (372 kbps)

Finally got to have a listen to your sample after a few other things, including a certain AAC 128 kbit test...

Without going into too much detail, I could also ABX this at 265, 320, and 400 with Wavpack lossy, high quality enabled, with and without joint stereo. Like you mention, it does seem to show the added noise quite readily, particularly at the lower bitrate end. There certainly are some types of music where the hiss can be more obvious, particularly high quality recordings where there is plenty of space in the music. With modern pop/rock music, the added noise tends to get masked by the music itself.

I imagine that this hiss may make Wavpack lossy less atractive to classical music listeners, where solo passages are more common and the individual instruments are usually more easily picked out when listening.

I didn't get a chance to try your sample with mpc. I'll take you on your word regarding the effect of the different encoders.

Den.

Thank you for testing. I should find a (very) low volume but not difficult sample, in order to test the noise level.

Well, this has been a very interesting sample!

First off, I am not able to reliably ABX this even at 265 kbps, much less hear a difference at 400! The quantization noise that WavPack is adding is way lower than the music (especially at 400 kbps!) and I was starting to think maybe you guys were playing games with me .

Then I noticed on the spectrum analysis that the sample drops off pretty steeply above 10 kHz, and above 15 kHz WavPack's quantization noise starts to peek through. I've seen this before and never thought too much of it because that area is pretty hard to hear (except at high volume) and I would have thought it would have been masked by the music down below. But then I remembered what you said den about having to listen at almost painful volumes to hear this hiss and when I took into consideration that my hearing only goes to about 14-15 kHz, it all started to make sense. This sample is a problem because of noise added in the 14-20 kHz range! (Hopefully you two can both confirm that you can actually hear in this range...)

This also explains why 1st-order noise shaping makes things worse. Like I suggested in another discussion with DickD, in this case noise shaping is moving the noise from an area where it's completely masked by the music to an area where it stands out all by itself, and ATH-style noise shaping would have made this even worse!

In the long term (i.e. WavPack 4.0), some sort of dynamic scheme that adjusts the noise shaping with the music might offer a significant improvement. For now, I have created an experimental WavPack.exe that has noise shaping continuously adjustable from -1.0 to +1.0, where positive values move the noise up in frequency and negative values move it down. The resulting lossy files are totally compatible with all existing WavPack players, but it is only possible to create "correction" files for the fixed -s0 and -s1 values (because the players don't know about the new methods).

Using a shaping argument of "-s-1", the quantization noise in this sample stays well below the audio all the way up to the top (even at 265 kbps). I would be very interested if this setting at the various bitrates improves the results for you guys.

Thanks a lot for this sample and your testing, BTW!!

The experimental encoder is here:

http://www.wavpack.com/wp397x.zip

bryant:
I haven't tested myself recently, but I know that when I was last tested (~4 years ago) I was hearing tones up to 20 KHz, although there was some drop off as I got closer to it. Before I get flamed by the dubious, the test was conducted by a specialist, as part of my employer's occupational health and safety hearing monitoring program. Such tests are required here in 'Oz under the Occupational Health and Safety Codes Of Practice regarding Noise and Hearing Protection, every 5 or 10 years if I recall correctly for industrial/mining/manufacturing companies.

Certainly what I hear is a high pitched hiss, not unlike that associated with tape recording hiss when the Dolby NR is switched off, but not as bad. I haven't ever had a good look at some of my samples regarding frequency distribution, but your explanation of the noise being prevalent at the high end, with no signal to mask it would certainly make sense to me.

You know how when you listen to high frequency/near ultrasonic tones, you can sometimes almost feel them, rather than hear them as distinct tones? I sometimes get the same with Wavpack lossy. Apart from noise I can actually hear (hiss) at high volumes, I can sometimes almost feel it at lower volumes.

Sorry to write so subjectively rather than objectively which goes against the grain with HA, but that's how it works for me.

Thanks for the experimental encoder. I will have a play at the end of this week, and let you know how I go.

Den.

Hmm, sounds a little "socialist" to me! When companies can save money by not bothering to protect their employees' hearing it helps them compete in the global economy. What good is hearing over 15 kHz if you don't have a job!?

Just a little update on my progress. I have got an algorithm working that does a pretty good job of adjusting that shaping parameter dynamically by comparing the amplitude of the source at very high frequencies to the noise that it is adding. As the noise it's adding starts to reach the amplitude of what's already there, it starts turning on the noise shaping to reduce it. It slows the encoding way down, but it might be a nice option when someone wanted the absolute highest possible quality in lossy mode and didn't care about a correction file.

Thanks again...

Funnily enough, it was a Labour Government that brought it in about 15 years ago, and they are the closest we have to a socialist party. They have now been out of government since '96. It caused a real fuss in industry when it came out, and pushed up industry insurance premiums for the employers, because suddenly they are now liable for an employees loss of hearing during their working life, including losses from outside causes, such as listening ot loud music etc...

Don't worry anout us competing in the global economy. Thanks to our hard work ethic and lateral thinking, we do just fine.



Sounds cool. An update on my progress. I've made files at 265, 320 and 400 with and without the -s-1 switch. A quick and dirty ABX test with the 265 sample showed a remarkable improvement, but I wasn't in my normal testing environment, so I'll be testing them properly later today and posting my results tonight sometime.

I don't suppose you'd consider posting your newer encoder for testing... B)


Den.

@Den so what do you think the key bitrate is? I mean where quality gain becomes minimal

@bryant

OK, just completed my ABX tests. >_<

At 265, using guru's sample and your experimental encoder with -s-1, first attempt yielded 11/15...
Tried again, after having a closer listen to both samples, and hit 14/15 twice in a row at high listening volume. Compare this with 15/15 with my first attempt using your current 3.97 release at 265. I was hearing some slight noise/distortion in the percussion instruments themselves, no doubt due to the noise being stored more in the lower-mid frequencies now. It is very subtle, but I could hear it with practice.

For this sample, this new approach is a major step forward, and significantly improves 265 kbit as a usable bitrate now.

At 320 kbit, experimental encoder with -s-1, tried three times, and got 8/15, 7/15, 9/15...

Had a quick play at 400, -s-1, and didn't bother going any further... It was transparent as far as I was concerned.

I had a look at the frequency graphs in Audacity, and the experimental encoder certainly appears to more closely match the original with the -s-1 noise shaping compared to Wavpack 3.97, as you mentioned in your post, not that graphs are everything...

I might have a play with your experimental encoder with -s-1 on some other samples of my own, but it would probably be of limited use, as it will depend on the frequency distribution of the sample, as to whether it is the right place in the spectrum to hide the noise. Your idea of checking the signal as you encode, and switching in noise shaping when appropriate seems like the way to go for me. I would think that the performance hit in terms of speed would be acceptable, as Wavpack lossy is one of the fastest lossy encoders at the moment anyway. B)

Keep up the excellent work!

Den.

@westgroveg

With the normal encoder (3.97) it is somewhere between 320 and 400 kbits for me. At 320, I can usually still ABX it due to the presence of background hiss, but only at high volumes. At 400, the hiss disappears most of the time, but the odd sample still shows it up. I must stress only at high volumes though, and it does not bother me, which is why I settle on 320 as a bit rate I can live with quality and storage space wise.

With the experimental encoder, I haven't played with it long enough, by early impressions are that it could be between 265 and 320, if the right frequencies are chosen to mask the noise in. B)

It also depends what you mean by quality too. Even at 265, it does not have the artifacts that feature with other lossy codecs, just some hiss. As you crank up the bitrate, the hiss gets softer, to a point where you can not hear it. This level seems to depend on the music, the listener and the equipment used. But even at 265 kbits, I still prefer the hiss to the artifacts from other codecs.

Den.

This means WPL @ 265 quality is comparable to MPC Q8 plus WPL is a better format for transcoding, sounds good for my purposes. I think I'll wait until the next release before I start using WPL. Thanks Den.

Hum... I can't remember many people asserting that they are hearing flaws or artifacts with musepack braindead. And these hypothetical artifacts are certainly very, very rare (I remember that I heard a small difference on the jump sample, after a considerable effort). On the other side, the hiss of wavepack lossy is more constant, less difficult to hear - maybe more pleasant that any other artifacts, if artifacts there are...

In all case, there are no proofs of the identity or similarities between musepack@260 kbps and wavepack lossy@260 kbps. I can't understand why you are claiming that.

I didn't say they where similar just (judging from the ABX tests I've seen by Den) that the quality level's are comparable

You just said it yourself,

They are not hypothetical if users can still hear artifacts at this level.

What I tried to say, is that on exceptional occasions (critical signal, more than 100 listening of the same second), musepack --braindead is ABXable for me. On the same listening condition (> 100 listening), wavepack is ABXable too, but not on rare moment as with mpc, but on most musical passage. Therefore, I can't conceive that musepack@230 and wavepack@230 are comparable in general use. Transparency level of musepack is much higher at this bitrate (again, according to my hearing). At 350-400 kbps, it's another thing, but at 230 kbps, wavepack lossy have limits that mpc doesn't have.

I've done a lot of q8 encoding & in practice MPC with BrainDead profile is closer to 275kbps