Am Sam, 09 Okt 1999 schrieben Sie:
> >
> > I made a short test with lame-3.32.
> >
> > With the new Takehiro routines the encoder speed has improved.
> > However, I noticed some high frequency artefacts that were not present
> > in the last three or four releases. I tested both with -b 160 -m s and
> > with 128kbps settings. The ringing artefacts are minute but noticeable. I
> > also used xmms with a highpass engaged and with stereo enhancement effect
> > plugin activated to ease my listening tests.
> >
> > I�m not sure where this comes from.
> > Do Takehiro's and Leonid's latest enhancements comply?
> >
> > Could somebody check the implementation of late Mr. Huffman's invention?
> >
>
> I think they are compatible: Before switching to Takehiro's routines,
> I checked them pretty carefully: I encoded castanets.wav with
> Takehiro's routines, but with the ISO bitcount formula. The bits
> saved by Takehiro's routines were then wasted. This procedure
> produced an mp3 file different than the original mp3, since different
> Huffman tables where used for the lossless compression, but the
> decoded wav files were identical.
>
> So I am pretty sure the only difference in output between 3.32 and 3.31
> is that there are *more* bits available per frame.
>
> Could you try using a larger normalization constant in the ATH formula?
> The ATH stuff is relatively new and will have its biggest effect
> on high frequencies.
>
> On line 216 quantize.c:
>
> ath -= 114; /* MDCT scaling. From tests by macik and MUS420 code */
>
> Something like ath -= 200 will completely disable it.
>
> Mark
>
I also get artefacts with lame 3.32 that weren't in lame 3.31.
With my -X5 VBR code I can get up to the cutting edge of lames
PSY model. And the new code produces on the Glockenspiel
SQAM files VERY audible scratchy artefacts, the old code not.
And the difference in file size for gspi35_1 is 188624 bytes (lame 3.32)
compared to 188716 bytes (lame 3.31). For harp40_1 the new code
produces even larger files, 199774 bytes (lame 3.32) compared to
199712 bytes (lame 3.31).
To the ATH problematic. After some earlier listening tests with the SQAM files
I would suggest to change line 216 in quantice.c as follows:
ath = ath/3.5 - 114;
This was considered above.
Robert
SQAM files:
http://sound.media.mit.edu/mpeg4/audio/sqam/
--
MP3 ENCODER mailing list ( http://geek.rcc.se/mp3encoder/ )
