IPB

Welcome Guest ( Log In | Register )

1.01j encoder vs 1.15u, listening test inside
guruboolez
post Jun 18 2005, 06:41
Post #1





Group: Members (Donating)
Posts: 3474
Joined: 7-November 01
From: Strasbourg (France)
Member No.: 420



Yesterday, I had some free time to play with various encoders. Like many people on this board (I guess) I’m still waiting on a major AAC release (Apple’s VBR encoder; major release of Nero AAC or at least the ‘green light’ for using the ‘fast’ and high quality encoder), and I have therefore decided to wait before starting a new personal multiformat comparison. This disappointing situation gives me an occasion to play with other encoders, especially musepack audio.

First, I’ve encoded many classical samples with latest encoder, with –quality 4 (--radio) profile. After two encouraging results on collective tests, musepack’s quality and reputation are now well-established at this bitrate (~135 kbps). I’ve experienced it myself, and I clearly admit that MPC at mid-bitrate performs really well… with some exceptions.
In reality, the same collective tests revealed serious issue with musepack, especially with classical music: bloated bitrate with some samples first (up to 180-210 kbps on harpsichord), and the opposite with some other samples, especially low-volume moments. With the sample named Debussy.wav (opera part: two people speaking/singing with a small orchestral accompaniment wreathed in recording noise), musepack performance was much worse than other competitors during Roberto’s 2nd 128 multiformat collective test (worse than WMA and atrac3!). This unusual low performance was probably correlated to a strange bitrate plunge (~100 kbps), not bad by itself (as long as it not hurts, lower bitrate are always welcome), but really questionable when quality is so strangely bad. Of course, Debussy.wav is only one sample, but it’s not a hard task for someone enjoying classical music to find numerous moments of low-volume including recording noise…

My own experience of musepack at mid-bitrate is not a big one, and a good half of it could be found on the first (and partially flawed) listening test I’ve published twenty months ago (here). In October 2003, musepack 1.14 performed equally with vorbis, slightly better than lame 3.90.3, but worse than WMA9Pro (winner) and QuickTime AAC. In other words, the relative performance of MPC –radio was not as enjoying than quality obtained with less specific music genre. Last but not least, with all progress made during this time by AAC (both Nero & Apple implementations – and faac too), lame, vorbis and maybe WMA(pro and only pro), musepack could seriously pretend to be the potential looser among all other modern audio format.

Yesterday, by listening quickly to various classical samples (I’m building a full library of very short samples, yet unfinished), I was simply amazed by poor performance of musepack at this bitrate on my music (I insist: the encoder performs differently and much better with everything else than the general category of ‘classical’). I’m now used to live with ABR encodings coming from LAME 3.97a, and I could swear with a great confidence that LAME is free of most issues audible with musepack: ringing, lifeless sound, and a typical artifact of mpc which reminds me for unknown reasons a washing machine tongue.gif). I have suspected one moment the PNS tool (enabled by default up to –quality 4.99) to have a bad impact with some of these samples, but a quick try (---pns 0 disables it) showed this algorithm to be innocent. Then, I have decided to try older versions of the encoder. I’m used to amass a plethora of encoders and keep them on my hard disk, but recently I’ve burned them all, and had the good idea to loose the disc (lost somewhere in the shambles I’m using as an apartment). I’ve only kept mppenc 1.01j (please, don’t ask me why!). And result obtained with this encoder is completely different: less ringing, lower artifacts… quality is much better with mppenc 1.01j --radio than with mppenc 1.15u --radio with classical music.

Few words about musepack –radio history (and correct me if I’m wrong). Frank Klemm worked on musepack at mid/low bitrate few time after he released mppenc 1.00 (it must be 1.02, 1.04 or 1.06), by introducing a tool called PNS (different beast IIRC from MPEG-4 AAC Perceptual Noise Substitution). As a consequence, the use of PNS lowered the bitrate of some MPC profiles (radio, thumb, telephone) without jeopardizing the output quality. It means that 1.01j --radio will systematically be weightier than 1.15u --radio (~15 kbps) and it implies that both settings are not directly comparable. The comparison I made is logically unfair. But it nevertheless reveals unexpected drawbacks of former musepack tunings, and (I hope so) will maybe open the door of further improvements smile.gif

As example, Debussy’s encodings (replaygained with high value) could be downloaded here:
1.01j
1.15u
Debussy (lossless)


Ok, and then? Now that you’ve read my awful literature, you’re probably disappointed to not see any table of results, ABX score or ANOVA analysis. No? It’s simply because I haven’t made a complete comparison between 1.01j and 1.15u at --radio profile… in order to make one at the famous --standard setting. It might sound totally absurd to compare a three-year old version of mppenc (released on May 2002) to the most recent version of the encoder (February 2005). But if tunings made three years ago hurts on several classical music samples at --radio setting, it is natural to speculate about possible regression with other profiles. But I don’t really like unwarranted speculation (probably as dangerous -if not more- than placebo), and I’d rather make a listening test in order to obtain results, whatever they are, than being suspicious or throwing doubt on current release of musepack. My biggest risk is to hear nothing (good or wrong) between both encoders. But of course, if I have created a long topic and smashed the record of grammatical mistake per square centimeter, it’s not to tell to everybody that I can’t hear something wrong between two encoders considered for several years as transparent to most people wink.gif

This post has been edited by guruboolez: Jun 18 2005, 06:56
Go to the top of the page
+Quote Post
 
Start new topic
Replies
Lefungus
post Jun 20 2005, 07:45
Post #2





Group: Members
Posts: 86
Joined: 10-November 02
Member No.: 3745



Does someone have the sources of mppenc from 1.01j to 1.14 ?
Because it'll be much more difficult if not impossible to see where the regression was without those biggrin.gif


--------------------
It's a 'Jump to Conclusions Mat'. You see, you have this mat, with different CONCLUSIONS written on it that you could JUMP TO.
Go to the top of the page
+Quote Post

Posts in this topic
- guruboolez   1.01j encoder vs 1.15u   Jun 18 2005, 06:41
- - guruboolez   • samples: 15 classical samples, used on sev...   Jun 18 2005, 06:42
- - Vertigo   Guruboolez, you could have saved us the trouble of...   Jun 18 2005, 08:31
- - NumLOCK   Just a remark - as Guruboolez said: replaygained w...   Jun 18 2005, 11:11
|- - Garf   QUOTE (NumLOCK @ Jun 18 2005, 12:11 PM)Just a...   Jun 18 2005, 11:13
|- - guruboolez   QUOTE (NumLOCK @ Jun 18 2005, 11:11 AM)Just a...   Jun 18 2005, 13:44
- - madoka@ex-sounds   BTW, where is 1.01j binary?   Jun 18 2005, 11:51
|- - guruboolez   QUOTE (madoka@ex-sounds @ Jun 18 2005, 1...   Jun 18 2005, 13:39
|- - madoka@ex-sounds   QUOTE (guruboolez @ Jun 18 2005, 04:39 AM)QUO...   Jun 18 2005, 15:09
- - Digisurfer   Keep up the good work guruboolez. I always find yo...   Jun 18 2005, 17:01
- - JeanLuc   I'm so glad that all my Musepack encodes have ...   Jun 18 2005, 19:58
- - CiTay   Certain replies split to this thread.   Jun 18 2005, 20:40
- - Gabriel   Does the current MPC encoder adjust its ATH level ...   Jun 19 2005, 09:32
- - Gecko   Very interestig test, Guruboolez, thank you! I...   Jun 19 2005, 23:48
|- - guruboolez   QUOTE (Gecko @ Jun 19 2005, 11:48 PM)I think ...   Jun 20 2005, 00:40
- - Lefungus   Does someone have the sources of mppenc from 1.01j...   Jun 20 2005, 07:45
- - shadowking   Wouldn't it be easier to start with 1.14 and t...   Jun 20 2005, 08:07
- - guruboolez   Lefungus> mppenc was closed source when 1.01j w...   Jun 20 2005, 10:00
- - Peter   Case mentioned having at least 1.14 sources some t...   Jun 20 2005, 10:37


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

 



RSS Lo-Fi Version Time is now: 21st April 2014 - 04:54