On 2013-11-05, Fons Adriaensen wrote:
I have often wondered if there was a way to confirm matrix encoding
by using a scope to display phase differences, but have found that
not to be the case
It should be possible to do that, but the process will be a bit more
complicated.
This has also been talked about before.
Calculate a series of FFTs on overlapping windows, over the entire
lenght of the recording, and look for components that have equal
magnitude in L and R.
On my part, I'd rather do a continuous (perhaps FFT-enabled, perhaps
frequency segregated) wideband Hilbert transform, and in way or another
try to bin+aggregate the phase relationships between the two carrier
channels, so as to build up an histogram of where the phase
relationships landed over time. I believe, over time, that would be the
easiest and most adaptable way of gathering the basic data, to then be
subjected to some prima facie Bayesian reasoning and/or algorithms
derived from machine learning from-sample.
I suspect support vector machines derived from
Vapnik-Cervonekis-complexity-limited low order Chebyshev-polynomials
might be useful in the latter, too: their inverse problem is the most
well behaved I know with regard to all kinds of noise and nonlinearity,
in these kinds of problems, and at the same time, if you can somehow
marry them to an efficient preprocessing and prebinning stage, maybe you
could even do realtime, adaptive recognition of the original analog
encoding.
But obviously, since I've never actually coded anything like this out or
seen the intermediate statistics, this must remain just a hint to
someone more willing than me to "actually talk the talk".
If they were panned to center using a normal stereo pan pot they
should be exactly in phase.
That, and the fact that in analogue material you can't really rely upon
the stuff even staying within the same encoding regime, is why I above
mentioned time spans. And why I believe this sort of blind decoding, if
tried, should be able to do mixed decoding and/or seamless transitions
from the start. Because, once you put in even a statistically optimal
recognizer, it'll be just half sure, half of the time, and will be
telling you lots of unexpected things in the middle of that
half-and-half.
Thus, the worst problem might not be to detect what you want or do not
want to see at all, and then just decode what is there. The real problem
might be to deal with continually and variously detecting something in
between, and how to decode that, then, without sounding hideous/silly
pretty much all of the time.
If the signal is UHJ encoded AMB there will be a phase difference of
around 35 degrees.
In fact, under some rather mundane assumptions, at least two channel UHJ
can even be automatically and reliably detected to the level of
resolving L from R. The same isn't true for e.g. Dolby MR, where the
left-right difference doesn't exist at a level deeper than simple 180
degree phase reversal.
If you find that consistently on all center front panned components
that would be strong evidence of UHJ encoding.
So, exactly that. At the same time, the accumulation over the Scheiber
sphere of phase-amplitude points, combined with a couple of rather
minimalistic a priories on what sound fields ought to look like in real
life, ought to be able to statistically differentiate between pretty
much all of the extant, untagged, analogue systems. MP/SQ/QS/UHJ/RM, all
that typical stuff I at least know about, and with a couple of tricks
over time, probably even the various extant versions of each of those
systems. In fact, using algorithms derived from the audio encoding lit
of the past ten years or so of mobile phone codec algorithms, you should
even be able to efficiently (i.e. in real time) compensate for static
delay differences between the channels, slowly varying same, quite a lot
of backround noise, and even certain kinds of speaker-microphone-like
rapidly varying angular distortion. In some regards even Doppler, even
if it's particularly nasty, as a non-shift-invariant phenomenon.
So, exactly that, and even more.
Looking at the complete signal it's probably impossible to decide if
any phase difference is significant or not, you need the 'logic
decoding' first.
The way I see it myself, as an amateur (and perhaps soon freelance
professional) audio DSP guy, as well as an amateur economist, is that
you shouldn't look at the instantaneous phase differentials as such.
Instead, you should recognize that there is a whole spectrum of
different timespans between none at all and hours, relevant to the
solution, all of them interworking the whole time.
Granted, I might be making this a bit more difficult from the start than
it has to be. But still, think about a typical audio feed from your
current Western television set. It does have commercial breaks in it,
no, which might contain totally fucked up audio wrt the audio you had in
the movie they just interrupted. No? So in principle, do you not have to
switch fluently and rather often between encodings, not only in kind,
but in time as well? I think you have to be able to do that.
Also, you have to be able to do that even in mixture, because in most
work I've heard of, "the stuff" was *not* mixed any single system. As I
recall, even such iconic things as the Star Wars soundtrack (almost)
always contained ad hoc elements which didn't utilize the underlying
Dolby encoding, but were placed as such onto the raw channels. If I
remember right, you'd have to be able to decode even time-variant
mixtures, from the start, in order to be useful in the real world. And,
you in fact *can* do that, at least in theory, but what I'm then saying
is that the methods you'd use to do that can't stay within the typical
single-rate LTI framework; the'll have to be multistage and multirate,
often nonlinear ones.
(And actually I might have to ask the group a couple of salient
questions here, in a short time, because it just might be I have to
utilize some ambisonic methods just the way I've described, on short
order. ;) )
--
Sampo Syreeni, aka decoy - [email protected], http://decoy.iki.fi/front
+358-40-3255353, 025E D175 ABE5 027C 9494 EEB0 E090 8BA9 0509 85C2
_______________________________________________
Sursound mailing list
[email protected]
https://mail.music.vt.edu/mailman/listinfo/sursound
