> Sounds like a good idea to me. But what options at what settings?
Presumably
> we want the settings with the best quality/speed tradeoff to be added at
> lower settings, with more expensive and/or less effective changes as -q
> increases. But some options have benefits that depend on other settings -
> for exmple, full huffman search might make a big difference at 32kbps, but
I
> doubt it's really noticable at 256kbps. Things like the more accurate MDCT
> quantization should give benefits at all bitrates.
Even if it would probably be smarter, I don't think we must turn different
option with the same -h x according to the bitrate. I think that it would
puzzle most users.
> BTW, does anyone have any idea what extra things MP3Enc does at its higher
> quality settings? Speed drops about five fold going from -qual 6 to -qual
7.
> The only quality setting listed in the docs that I would think could cause
> such a huge speed decrease is "many outer loops", but I've never found a
> perceptible quality decrease in dropping from -qual 7 to -qual 6. -qual 8
> and above enables full huffman search, but at the bitrates I typically use
> (192 or 224kbps) that's (again) probably not very significant.
According to the mp3enc doc, there are only 4 things affected by the -q x
setting, so there are obviously some -q x wich are identical (". In future
versions of the encoder, more features might be controlled by this switch").
The 4 things are:
A)Use a high-quality time domain filter instead of fast MDCT ("Soft
time-domain filtering")
B)Use the best sample rate without regard to filter running time. Adapting
to this sample rate might use CPU-intensive filtering ("Best match sampling
rate")
C)Find the best huffman code book possible to encode the spectrum of each
frame. A few percent bits can be saved in each frame, available for higher
quality in following frames. ("Full huffman search")
D)Shape the quantization noise very carefully ("Many outer loops")
According to the doc, A seems to be enabled starting with -q 1.
According to Mat, D seems to be anabled with -q 7 and C with -q 8
Here's my suggestion for LAME:
*-h 0
*-h 1
*-h 2: m/s masking thresholds
*-h 3
*-h 4: more accurate quantization (quantize_xrpow)
*-h 5
*-h 6
*-h 7
*-h 8: full Huffman search (best_huffman_divide)
*-h 9: subblock gain code
Regards,
Gabriel Bouvigne - France
[EMAIL PROTECTED]
icq: 12138873
MP3' Tech: www.mp3tech.org
(temporary down, alternative url is http://mp3tech.tsx.org )
--
MP3 ENCODER mailing list ( http://geek.rcc.se/mp3encoder/ )
