[R] Computing fundamental harmonics from a periodogram

[hadley wickham]

On 4/11/07, Uwe Ligges <ligges at statistik.uni-dortmund.de> wrote:
>
>
> hadley wickham wrote:
> > On 4/11/07, Uwe Ligges <ligges at statistik.uni-dortmund.de> wrote:
> >>
> >>
> >> hadley wickham wrote:
> >> > Dear all,
> >> >
> >> > I'm trying to finding the fundamental harmonics (ie. peaks in a
> >> > periodogram) from a time series (extracted from an mp3).  For example,
> >> > if I look at
> >> >
> >> > spectrum(fdeaths, spans = c(3,3))
> >>
> >>
> >> A heuristical procedure for finding fundamentals of a monophonic sound
> >> can be found in package tuneR. For a short example how to use it, type:
> >>
> >> install.packages("tuneR")
> >> library(tuneR)
> >> ?tuneR
> >>
> >>
> >> For polyphonic sound, I do not know any method that is both sufficiently
> >> fast (i.e. calculates results for a 1 minute sound in less than a day)
> >> and accurate.
> >
> > Thanks Uwe, I have been using tuneR already, but I have been unable to
> > extract the peaks out of the periodogram.  I'm trying to follow along
> >
> > @book{wieczorkowska:2005,
> >     Author = {Wieczorkowska, Alicja and Synak, Piotr and Lewis, Rory and
> > Ra{\AA}, Zbigniew W.},
> >     Publisher = {Springer},
> >     Title = {Extracting Emotions from Music Data},
> >     Year = {2005}}
> > http://www.springerlink.com/index/LRC0995XXL5M12X4.pdf
> >
> > but need the harmonics to calculcate most of the variables they are using.
>
> Interesting! Printer is already working to make a hardcopy.
>
> To your problem:
> Even voting which harmonics are for which fundamental is difficult,
> since for polyphony, the problem is not identifiable (i.e. there is no
> general unique solution; e.g. quints, octaves etc. do share some
> partials) ...
>
> I know of some heuristics by a physician in our university who has
> invented an algorithm (still unpublished) that extracts the partials
> (and of course has the problems mentioned above for polyphony) on its
> way to another target. We (i.e. a phd student rather than me) are in the
> progress of trying to implement the algorithm - which appears to be
> rather complicated  ...

Thanks Uwe, I had hoped that there would be something off-the-shelf I
could use.  It's not a big deal, I was just trying to generate an
interesting data set for a visualisation class that I'm teaching.  The
very simple statistics that I had been extracting with tuneR (energy,
variance, pureFF, etc) don't discriminate between difference genres of
music (let alone artistcs) at all well.

Hadley