Musepack vs other codecs

[2M2B]

With how in 2021 DAP's(Fiio m6) and Iphone/smartphone have 3rd party music apps. It is shocking at how this format based on MP2/subband can be transparent at 160kbps VBR with no killer samples, It even beats many MDCT codecs on Noise/Electronic/Harsh classical(In Slaughter Natives) at a average rate of 96 ~ 190kbps, While much harder stuff will use bit rates of 330 ~ 1300kbps. Yup, It possible to have a file be 1mbit at It's 160kbps VBR setting already have some tracks with sections that reach 780 ~ 890kbps.

 

[q3cpma]

It is good and decode fast, but I would still use Opus. The comparison with Vorbis/AAC is more interesting. My biggest problem with Musepack is that it's abandonware and that it doesn't use Vorbis comments which are way better for the user/programmer.

 

[2M2B]

It is good and decode fast, but I would still use Opus. The comparison with Vorbis/AAC is more interesting. My biggest problem with Musepack is that it's abandonware and that it doesn't use Vorbis comments which are way better for the user/programmer.

It finalised codec the 192kbps blind test done on Hydrogen Audio show it transparent at that bit rate. Every app I used that plays music has .mpc support. Vorbis/AAC are patchy messes compared to 160kbps MPC & Opus, V2 Lame holds up better despite the pre echo problems.

Vorbis fit's abandonware term better since the dev never fixed the puffy wind artefact if the any noisy transients and needs 384kbps VBR(Q9.2) to stop it. It just small patches to stop it bugging out, no massive quality fixes. It why I stick with MP3 on gear with limited codec support.

 

[UKPI]

It is shocking at how this format based on MP2/subband can be transparent at 160kbps VBR with no killer samples,

Really?

Spoiler: ABX1

foo_abx 2.0.6d report
foobar2000 v1.6.6
2021-05-22 23:39:15

File A: 41_30sec.flac
SHA1: 264e5f7ef182235b1700a22339e6736d706f2257
File B: --41_30sec.mpc
SHA1: bb3ce4b06a71e5c8ef9e5101f2b0625f3bc45782

Used DSPs:
Equalizer, IIR Filter, IIR Filter, IIR Filter, Meier Crossfeed, Scale, Resampler (SoX), Advanced Limiter

Output:
WASAPI (push) : 헤드폰(Meizu HiFi DAC Headphone Amplifier PRO), 32-bit
Crossfading: NO

23:39:15 : Test started.
23:39:50 : 01/01
23:40:36 : 02/02
23:41:26 : 02/03
23:42:40 : 03/04
23:43:16 : 04/05
23:43:37 : 05/06
23:45:03 : 05/07
23:45:36 : 06/08
23:46:00 : 07/09
23:46:59 : 08/10
23:47:20 : 09/11
23:48:06 : 10/12
23:48:06 : Test finished.

----------
Total: 10/12
p-value: 0.0193 (1.93%)

-- signature --
c4eecc88049f5cab31c15b6de4fb186abf0ca5f9

Spoiler: ABX2

foo_abx 2.0.6d report
foobar2000 v1.6.6
2021-05-23 00:25:07

File A: KSJ.flac
SHA1: 519023bacc5d955ab6aefc7ca808dbbee9d439d3
File B: --KSJ.mpc
SHA1: 53b0a1ae8cf7e84f17bc8b1e5d09b1b43659f4c1

Used DSPs:
Equalizer, IIR Filter, IIR Filter, IIR Filter, Meier Crossfeed, Scale, Resampler (SoX), Advanced Limiter

Output:
WASAPI (push) : 헤드폰(Meizu HiFi DAC Headphone Amplifier PRO), 32-bit
Crossfading: NO

00:25:07 : Test started.
00:26:18 : 01/01
00:26:39 : 02/02
00:27:00 : 03/03
00:27:36 : 04/04
00:28:24 : 05/05
00:28:34 : 06/06
00:28:45 : 07/07
00:29:54 : 08/08
00:30:19 : 09/09
00:30:42 : 10/10
00:31:15 : 11/11
00:31:25 : 12/12
00:31:25 : Test finished.

----------
Total: 12/12
p-value: 0.0002 (0.02%)

-- signature --
15d38c5dc8521c1b537382dc25d69db0a4747e66

Spoiler: ABX3

foo_abx 2.0.6d report
foobar2000 v1.6.6
2021-05-24 02:41:05

File A: velvet.flac
SHA1: ea3e7e20c456376884f6515ccbd05db753f91740
File B: --velvet1.mpc
SHA1: 5f919c4e987aedb47fb4aac6a4bcc49f8ebf4616

Used DSPs:
Equalizer, IIR Filter, IIR Filter, IIR Filter, Meier Crossfeed, Scale, Resampler (SoX), Advanced Limiter

Output:
WASAPI (push) : 헤드폰(Meizu HiFi DAC Headphone Amplifier PRO), 32-bit
Crossfading: NO

02:41:05 : Test started.
02:41:24 : 01/01
02:41:45 : 02/02
02:42:00 : 03/03
02:42:28 : 04/04
02:43:27 : 05/05
02:43:54 : 06/06
02:44:27 : 07/07
02:44:43 : 08/08
02:45:16 : 08/09
02:45:42 : 09/10
02:45:56 : 10/11
02:46:10 : 10/12
02:46:10 : Test finished.

----------
Total: 10/12
p-value: 0.0193 (1.93%)

-- signature --
1f7aac5faf2e025bbea89463a97ccf994e454aa6

Not really.

The artifacts of --velvet1.mpc and --41_30sec.mpc were subtle (hence the errors in ABX1 and 3) but they were there at the standard setting. KSJ.flac was the killer sample. To me, the latest version of Fraunhofer FDK encoder was transparent on all three samples at comparable bitrates. (VBR mode 5, 17kHz cutoff option was used.) I personally won't judge Musepack to be better than other conventional codecs like FDK AAC and aoTuv-tuned Vorbis at bitrates over 160kbps.

I've uploaded the smaller files here. The rest can be found at: https://www.rarewares.org/test_samples/

While much harder stuff will use bit rates of 330 ~ 1300kbps. Yup, It possible to have a file be 1mbit at It's 160kbps VBR setting already have some tracks with sections that reach 780 ~ 890kbps.

The bitrate of AAC can reach up to 529kbps at 44.1kHz. Vorbis doesn't even have that limit. It is possible to create files with some parts reaching 500+kbps, but encoders generally don't do that because it is a waste of space/bandwidth. There's no reason to throw in more bits when the quality improvement is minimal.

Vorbis fit's abandonware term better since the dev never fixed the puffy wind artefact if the any noisy transients and needs 384kbps VBR(Q9.2) to stop it.

Were blind tests (ABX, ABC/Hr) performed to reach that conclusion? I don't think Musepack will receive new fixes for the artifacts that I've encountered in the near future.

 

[UKPI]

Never said the codec 100% foolproof

I just refuted the claim that Musepack is transparent and AAC/Vorbis are a patchy mess compared to it at a bitrate of 160+ kbps. If that claim was close to the truth, it would have been much harder for me to find multiple samples with audible distortions.

did you try It at 256kbps before dismissing the codec?

Even an MP2 encoder which uses the psychoacoustic model of the ISO reference implementation of MP3 is very close to being transparent at that bitrate. (I haven't found a sample that causes audible distortions for it yet. Someone else with better hearing and more interest in it might.) Achieving near transparency at 256kbps isn't that special.

Of course, labeling Musepack as an inferior codec only with the data above would be as misleading as calling it a superior codec. That's why I just wrote that Musepack doesn't seem to be better (which is different from being worse) than AAC or Vorbis at high bitrates.

Vorbis sucks with as limited support as MPC.

That was one of the big drawbacks of Vorbis back in the 2000s. At least it was a free codec that could potentially threaten the position of MP3/AAC and played at least some part in keeping the licensing fees low.

 

[2M2B]

I just refuted the claim that Musepack is transparent and AAC/Vorbis are a patchy mess compared to it at a bitrate of 160+ kbps. If that claim was close to the truth, it would have been much harder for me to find multiple samples with audible distortions.

I have done my tests under double blind where I have no clue what codec I'm doing. After I thought I could tell Vomir & The Rita at MPC Q5(88 ~ 132kbps) vs AAC/Vorbis at 160kbps with high/low anchors. I just did one few days ago where I thought I picked 256kbps MPC then found after It was done I picked the 128kbps as transparent copy. The puffy noise Vorbis has on anything with crunchy noise textures, Is from micro attack problems. The baron chamber which has noise burst near the end Vorbis craps out while MPC/AAC at 160k & MP3 at 192k sound fine. I've tried those samples before on HA since many were used to tune the LAME encoder in the past.

Dead body collection's - trypophobia shows MPC's VBR system is much better tuned, It pumps out a 300kbps file. While Opus needs 400kbps, AAC breaks at 500k, Vorbis needs 500kb/s, and maybe Lame at V0 with the --allshort & a 18KHz lowpass. This is what many newer Video codecs do in VBR mode on hard samples, Like let say on easy to moderate stuff it stays at 7mbit at 1080p but when thing get chaotic it will up the bitrate to 22mbit if needed.

 

[2M2B]

Even an MP2 encoder which uses the psychoacoustic model of the ISO reference implementation of MP3 is very close to being transparent at that bitrate. (I haven't found a sample that causes audible distortions for it yet. Someone else with better hearing and more interest in it might.) Achieving near transparency at 256kbps isn't that special.

Musepack is proving what could've been if FHG pushed MP2 instead of MP3. Because I remember when that company said 128k MP3 >> 192k MP2, But then when quiet when MP3 could not cope with pre echo even at 320kbps. Even If 160k MPC can have slight artifacts upping the track to 256kbps can fix which can not be said for the MDCT based ones which is why HA never liked MPC in high bitrate tests, It always won without needing do much blind testing. As my wall of text above phantom artifacts is a thing since MPC artifacts are totally different to what breaks MDCT/Hybrid codecs.

 

[digitalfrost]

Musepack was one of the best codecs back in the day, see also https://en.wikipedia.org/wiki/Musepack#Test_results.

I have a ton of music in MPC format. Today I wouldn't use it anymore. First of all, there is no support with most devices that are not a PC, and even on PCs not many audio players support it. Also, the other codecs have gotten so good, I'd just use LAME MP3 and not have to worry about it anymore.

Also, storage costs have gone down so much if I wanted transparency I'd just use FLAC.

 

[2M2B]

First of all, there is no support with most devices that are not a PC, and even on PCs not many audio players support it

So using Neutron player and Foobar mobile don't count?. The Fiio M6 uses Android which you could side load Foobar mobile or Neutron on it. It not 2009 any more.

Also, the other codecs have gotten so good, I'd just use LAME MP3 and not have to worry about it anymore.

V2 Lame MP3 performs on par with other 4 codecs at 128kbps. So no wonder people can pick out 256/320 MP3 much easier than the other 4 codecs at 192kbps. But It is weird how Vorbis the worst performing of the 4 codecs to the right, can be worse than MP3 on most samples not used in that test.

 

[UKPI]

I have done my tests under double blind where I have no clue what codec I'm doing. After I thought I could tell Vomir & The Rita at MPC Q5(88 ~ 132kbps) vs AAC/Vorbis at 160kbps with high/low anchors. I just did one few days ago where I thought I picked 256kbps MPC then found after It was done I picked the 128kbps as transparent copy. The puffy noise Vorbis has on anything with crunchy noise textures, Is from micro attack problems. The baron chamber which has noise burst near the end Vorbis craps out while MPC/AAC at 160k & MP3 at 192k sound fine. I've tried those samples before on HA since many were used to tune the LAME encoder in the past.

Same can be said with any well-designed encoder for post-MP3 formats with different samples from people with different sensitivities and preferences for artifacts. That's not unique to Musepack.

Musepack is proving what could've been if FHG pushed MP2 instead of MP3. Because I remember when that company said 128k MP3 >> 192k MP2, But then when quiet when MP3 could not cope with pre echo even at 320kbps. Even If 160k MPC can have slight artifacts upping the track to 256kbps can fix which can not be said for the MDCT based ones

Just like a lot of companies, Fraunhofer said many things. There was a claim of 128kbps mp3 reaching near CD quality in the instruction manual of the demo version of mp3enc released by them in 1998. Those marketing materials from nearly two decades ago are irrelevant in deciding which modern codec is the better one.

Also, techniques that reduce or hide pre-echo are implemented in MDCT based codecs. AAC LC uses TNS (moving pre-echoes closer in time domain to the transients so that those can be masked better) and uses smaller windows for transients compared to MP3. Vorbis has a more flexible set of window sizes from 64 to 8192 which are more flexible than AAC LC and MP3.

So using Neutron player and Foobar mobile don't count?. The Fiio M6 uses Android which you could side load Foobar mobile or Neutron on it. It not 2009 any more.

V2 Lame MP3 performs on par with other 4 codecs at 128kbps. So no wonder people can pick out 256/320 MP3 much easier than the other 4 codecs at 192kbps. But It is weird how Vorbis the worst performing of the 4 codecs to the right, can be worse than MP3 on most samples not used in that test.

Yes, I've seen that test. That result comes from here. Hard to draw any conclusion from statistically insignificant results except MP3 from a single person.

To reiterate, I'm not claiming that Musepack is inferior. I'm just claiming that there is not enough evidence to judge Musepack to be superior to other codecs at higher bitrates. To verify that, a high bitrate multi-codec listening test involving multiple listeners is necessary. Without it, there are only speculations.

 

[2M2B]

Yes, I've seen that test. That result comes from here. Hard to draw any conclusion from statistically insignificant results except MP3 from a single person.

That because when HA did a large DBT with MPC at 128kbps before 2006. They got mad that MPC was 4.4 and AAC/Vorbis were 4.1, I think they did that to avoid MPC at 192kbps being a 5.0 and the rest(AAC, Opus, Vorbis) being 4,6. Since I've posted samples where AAC, Vorbis, Opus need 480kbps VBR were ignored, But then act like MPC being 4.8 is massive issue.

 

[2M2B]

foo_abx 2.0.6c report
foobar2000 v1.6.7 beta 15
2021-07-17 08:39:38

File A: 41_30sec.wv
SHA1: 9b7624f6a86bbd801deca11a5c33317268075fa2
File B: velvet.mpc
SHA1: 5947d117f79f26076b8de1072667af9c5670d01a

Output:
Default : Primary Sound Driver
Crossfading: NO

08:39:38 : Test started.
08:40:13 : 01/01
08:40:45 : 01/02
08:41:14 : 01/03
08:41:43 : 02/04
08:41:59 : 03/05
08:42:28 : 04/06
08:42:47 : 05/07
08:43:02 : 05/08
08:43:19 : 05/09
08:43:35 : 05/10
08:43:55 : 06/11
08:44:17 : 07/12
08:44:17 : Test finished.

----------
Total: 7/12
p-value: 0.3872 (38.72%)

-- signature --
72ff520a0bfc1d2de188bc154cc064c8fef5b003

Pretty much transparent blows away MP3 on HF handling & avoid the fake HF that AAC/Vorbis do to hide there only 16KHz.

foo_abx 2.0.6c report
foobar2000 v1.6.7 beta 15
2021-07-17 08:52:20

File A: KSJ.flac
SHA1: 519023bacc5d955ab6aefc7ca808dbbee9d439d3
File B: KSJ.mpc
SHA1: 53b0a1ae8cf7e84f17bc8b1e5d09b1b43659f4c1

Output:
Default : Primary Sound Driver
Crossfading: NO

08:52:20 : Test started.
08:52:57 : 01/01
08:53:11 : 02/02
08:53:24 : 02/03
08:53:37 : 02/04
08:53:51 : 03/05
08:54:05 : 03/06
08:54:17 : 04/07
08:54:30 : 05/08
08:54:42 : 05/09
08:54:55 : 06/10
08:55:11 : 06/11
08:55:23 : 07/12
08:55:23 : Test finished.

----------
Total: 7/12
p-value: 0.3872 (38.72%)

-- signature --
e502ec59c58f391f51ae1fa624c2b248df41fe32

Couldn't hear anything, Still transparent.