[prev in list] [next in list] [prev in thread] [next in thread] 

List:       mp3encoder
Subject:    RE: [MP3 ENCODER] Intensity stereo
From:       "M. Alexander Broadhead" <alex.broadhead () clearband ! com>
Date:       2001-11-15 16:12:39
Message-ID: 000a01c16df0$59ebd190$1a09010a () clearband ! net
[Download RAW message or body]

Howdy Gabriel,

> > > I was naively thinking that determining the right intensity
> > > position for
> > > each subband could also be a problem.
> >
> > No, this value is completely determined by the signal.
> It's an expensive
> > computation, but it's not complicated.
>
> How do you determine the position?

>From ISO/IEC 11172-3, Appendix G, p.144:

 is_pos(sb) = NINT((12/PI)*arctan(sqrt(L_Energy(sb)/R_Energy(sb))))

Like I said, expensive (given the sqrt() and arctan(), as well as a divide),
but straightforward.

> > Hmmm...  I hadn't thought of trying to guess the threshold from the
> signal.
> > The idea clearly has merit; after all, that's how mid/side
> works.  I was
> > just planning to check post facto, i.e. apply each level of
> intensity, and
> > see if it's an improvement or not.  (Which is how
> Layer-I,II do joint.)
>
> You do encoding for each configuration? Well, this method is
> more likely to
> give you the best results, but it's sure it will be very very slow.
> Perhaps (in a given sfb) the following could work:
> Compute the average left and right power. Now, compute the
> ratio between
> left and right power. This will be our reference.
> Now, let's do a loop over each frequency band of this sfb.
> For each freq,
> compute the ratio between left and right. Compute the number
> of freqs that
> have a ratio that is more than 10% away of the average ratio.
> If less than
> 15% of freqs in this sfb are too far away (more than 10%) of
> the average
> ratio, then this sfb is qualified for is. Of course 10 and
> 15% are arbitrary
> and need to be determined by experiment.

I'm just not sure deviation of the panning from the average panning will
correlate with quality loss.  It's probably worth running some experiments,
though.

My (brute force) method got around that by just computing all the
permutations and picking the best one.  An obvious next pass on my method
would be to see if the quality generally has a single maximum about which
the curve of quality is monotonically decreasing; if so, divide & conquer
would search the entire range of 21 possibilities in only 5 passes (and so
would only be 5 times slower, which would generally stink, but not be
unusable).

> > Also, the concept of intensity is based on high-frequency
> signals not
> > needing stereo information.  You'd probably want to do some
> research to
> see
> > whether similarity of shapes was actually relevant, and, if
> not, how low
> in
> > frequency you can progress before it is.
>
> Yes, you're right for very high sfbs. In sfb21 (as an
> exemple), it's very
> likely to be noise-like, and the left and right shapes are
> likely to be
> decorrelated.
> But I think my upper idea could be of use. Perhaps only in a given sfb
> range?

I really wonder if no one else has done this sort of work before...  I'm
guessing all the relevant research is probably proprietary, though.

Alex



[prev in list] [next in list] [prev in thread] [next in thread]