> 
> Hi,all,
>    I have some questions to discuss with you:
>    1)i used some sine wave whose frequencies are 997,1k,etc to encode with
> lame mp3 encoder, but i found 997.mp3 is not as good as 1k.mp3. then i use
> soundforge to check theirs decoding wav, and i found some errors in decoding
> 997.wav, why?

Ths sounds like a bug.  But the first thing to do before reporting
such a bug is to decode the mp3 files with 'lame --decode'.  A lot of
decoders out there have bugs too.  


>    2)i have tried to encode without Psymodel, and i tested the result using
> a sine wave, it sounds well and it's SNR is approx equal to the old. And i
> also tried to encode music, they sounds well too, so i want to do fixed
> point encoding without Psymodel, but whether does it do well for all music?
>     3)And,you know that the results of subband systhesis are also greater
> than 2^31or less than -2^31, and almost all data of the table enwindow[] is
> divided by 2.384e-06, so, to lessen the results, i change 2.384e-06 to
> 2.384e-04 first, and multipled win[4][36] with 1e2, the results can be
> represent with int! But the precision is not enough, and the SNR of final
> mp3 encoding with this method isn't as good as old. Before, i calc subband
> systhesis using exponent+mantissa, but it runs slowly. How to maintain
> precision and the results don't exceed int? What about your opinion?
> 
> best regards
> anne
> 

You should be able to get pretty good results without the maskings
from the psymodel, but you probably still need pre-echo detection and
some default masking values.  All the additional masking computations
is probably only a 10-20% improvement.  It's necessary if you want
to compete on quality, but not necessary if you are just putting
together an embedded app to sell to people who dont know any better :-)

As for the precision problems, with carefull scaling (and you may have
to rescale many times for different parts of the algorithm) I think
you should be okay with 32 bit integers.  But it will require a
detailed understanding of every algorithm.  I have no idea if
commercial embedded encoders are using all integers, or software
floating point - what ever they are doing it is probably a closely
guarded secret!


Mark




















_______________________________________________
mp3encoder mailing list
[EMAIL PROTECTED]
http://minnie.tuhs.org/mailman/listinfo/mp3encoder

Reply via email to