HydrogenAudio

Hydrogenaudio Forum => Listening Tests => Topic started by: legg on 2005-06-28 01:46:00

Title: Volunteers for ABC/HR listening tests needed
Post by: legg on 2005-06-28 01:46:00
Only ABC/HR, no ABX needed, this should make tests faster and give an indication of the expected quality rather than a boolean indicator of transparency.

I implemented a transform codec as my thesis project, the work is finished and all I need is some subjective evaluation of the quality expected. I'm comparing it against LAME @ 256kbps and Nero LC AAC @ 256kbps.

The most predominant artifacts are pre-echo (no adaptive block), distortion on highly tonal signals and right before silence, aka fade out (I suspect the bit allocation routine needs improvements).

Anyway, for those interested in helping, I'm using ff123's ABC/HR program, I was adviced later by guruboolez to use ABC/HR for Java, but I was already gathering information on the other program.

The download is 100MB, all reference files are losslessly compressed using FLAC, I made a .bat file to decompress them all to WAV as required by the program, running decode.bat is all you need (as flac.exe is also included *gasp*). The ABC/HR program is also included (I hope ff123 does not have a problem with this), and the tests are already set up, all you need to do is open the test files (the text files on the root directory) and start testing.

The URL for the download is here (http://lsc.fie.umich.mx/~legg/tesis/ABCHR.zip)

Thanks in advance, any help will be much greatly appreciated and cited in the document.

Edit: to add the fade out artifact, which I forgot to mention.
Title: Volunteers for ABC/HR listening tests needed
Post by: rjamorim on 2005-06-28 01:57:14
Quote
(I hope ff123 does not have a problem with this)[a href="index.php?act=findpost&pid=309376"][{POST_SNAPBACK}][/a]


I'm sure he won't, otherwise he wouldn't have licensed it under the LGPL
Title: Volunteers for ABC/HR listening tests needed
Post by: SebastianG on 2005-06-28 13:06:28
Did your codec also operate at 256 kbps ?
Care to explain some details about the inner workings of your codec ? For example: What kind of transform did you use / Does the transform adapt to the signal / How's quantization done / How's lossless coding done ? (You don't have to answer. It's just me being curious)

Unfortunately 100 MB is a little too much for a guy like me (56k dial-up connection).

Sebi
Title: Volunteers for ABC/HR listening tests needed
Post by: Defsac on 2005-06-28 13:25:13
I got a CRC error on newagelegg251.32.flac. It's possible it's a corrupt download though.

Quote
Extracting newagelegg251.32.flac
bad CRC 8b756782  (should be 8e4c7b68)
Title: Volunteers for ABC/HR listening tests needed
Post by: legg on 2005-06-28 16:23:19
SebastianG, The codec is pure VBR, on some test files the ABR is less than 256 (on fingersnaps it is 125.92) on some others it grows higher (312kbps on rock), since the rate is so variable, I considered that a fair compare would be against CBR 256kbps since I consider AAC y MP3 far better tuned than my own.

The transform used is MDCT, it does not adapt to the signal (hence the pre-echo), quantization is done following your suggestion of using powers of 2. The lossless coding part is done using arithmetic coding.

I do not do lowpass, this might be a mistake on my behalf but while doing the ABC/HR tests I noticed that some of them sounded "low-passed" and it corresponded to the AAC files (i'm talking classic and violins1).

Defsac, I checked the files and the integrity is fine, I'm uploading the file you need anyway, file here (http://lsc.fie.umich.mx/~legg/tesis/newagelegg251.32.flac)

Thanks for the help.