Xref: utzoo rec.audio:8759 sci.electronics:4167
Path: utzoo!utgpu!water!watmath!clyde!att!osu-cis!tut.cis.ohio-state.edu!mailrus!nrl-cmf!cmcl2!rna!dan
From: dan@rna.UUCP (Dan Ts'o)
Newsgroups: rec.audio,sci.electronics
Subject: Re: Extracting Mono from Stereo
Keywords: mono stereo signal processing
Message-ID: <287@rna.UUCP>
Date: 28 Oct 88 01:13:32 GMT
References: <341@ivucsb.UUCP> <343@ivucsb.UUCP> <3836@ncrcae.Columbia.NCR.COM>
Reply-To: dan@rna.UUCP (Dan Ts'o)
Organization: Rockefeller University Neurobiology
Lines: 33

)In article <343@ivucsb.UUCP> todd@ivucsb.UUCP (Todd Day <todd>) writes:
)I have a very interesting signal processing problem for those signal
)processing gurus.
)I want to use a DSP chip to extract mono from stereo.  Does anybody
)know of any algorithms to accomplish this or can anyone direct me to
)any literature that might describe this?
)I've looked into the math behind this.  It is not a simple as it first
)appears!  Now, it's very, very easy to get rid of mono from stereo...
)just subtract left from right, and the mono disappears.  However, you
)can't just subtract this (R-L) from (R+L) to get only mono.  What
)you always end up with is some combination of the mono plus one of
)the other channels (2M - 2R, etc.)

	Yes, an interesting problem -- one that I thought about a little
about a year ago, and I came to the same conclusions you did. As someone
else pointed out, you are trying to solve for three variables with two
equations.
	The auditory system can detect delays of onset in the 100us
range (it has "cute" neural circuits that compare the onset times
between sounds arriving at each of the two ears and fires with specific
delay times in the 100us resolution range). This type of processing,
combined with bandpass-like frequency response properties would allow
the auditory system to compare arrival times over a wide range of the
auditory spectrum.
	I should go ask some DSP friends, but perhaps some form of
cross-correlation processing would work. The output of the cross-correlation
should reflect the similarity between the R and L. Of course you would want
to restrict the time scale looked at to avoid including repeated passages in
the music. Unfortunately, you then must work backwards to somehow relate
the correlation to the signals responsible...hmmm... This problem seems
analogous to the frequency domain idea of sampling the spectrum with
bandpass filters or FFT's, determining the similarities, but then needing to
work back to the actual common mode portions of the music...