On 2012-03-07, Fons Adriaensen wrote:
That is more or less what I had in mind. Split up the signal into 'grains' in the time/frequency domain, identify those that are likely direct sounds, and build statistics of their L/R magnitude ratio and phase difference. Match these again encoding hypotheses.
As I said before, personally I'd like to start simple. Just apply a couple of Hilbert transformers to two encoded signals, giving four signals, then normalize the total vectorial amplitude/modulus away, leaving three signals. Then you're left with what is essentially a Scheiber sphere. Accumulate whatever happens on it, and utilize some version of Bayes's rule to discriminate between the different the different distributions painted on it.
I'm not certain this would work. But done right, even the encoding side noise ought to paint a clear locus on the sphere, which, when accumulated long enough, ought to reveal the encoding. And this sort of thing doesn't require any source separation machinery, so it's dead simple, and probably quite robust. If it works well.
-- Sampo Syreeni, aka decoy - [email protected], http://decoy.iki.fi/front +358-50-5756111, 025E D175 ABE5 027C 9494 EEB0 E090 8BA9 0509 85C2 _______________________________________________ Sursound mailing list [email protected] https://mail.music.vt.edu/mailman/listinfo/sursound
