Re: [Rhythmbox-devel] In the Mood plugin thoughts

[Charlotte Curtis <c f curtis gmail com>]

Hey guys!

I'm the "In the Mood" plugin writer, which you might have noticed
hasn't received any attention at all since SoC ended and I started
grad school (for which I apologize - it's not nice to be in the SoC
"failure" category, but somehow school takes up a lot of time).  For
that matter, I apparently haven't been reading this mailing list a
whole lot either, but at least I'm only a few days behind on this
discussion.

On Sun, Feb 8, 2009 at 9:04 AM, eric casteleijn <thisfred gmail com> wrote:
>
> > I did some snooping around, and it seems that there hasn't been much
> > development around playlist generation based on tonality.  Most of the
> > work is based on timbre.  If anyone could shed some light on the reasons
> > for that, I'm all ears.  Anyway, I think that a playlist generator based
> > on tonality and tempo would be really cool.  I will have to continue
> > researching to determine the feasibility, however.  I know the Serato
> > Scratch DJ software implements some reliable BPM and key analysis
> > features, so it must be possible.

I agree, tempo and tonality would be awesome.  My original plan for In
the Mood had adjustable song-matching controls, so that you could
specify an automatic playlist based primarily on tempo, timbre, etc.
I instead took the lazy approach of using an already developed
algorithm for feature extraction and matching (the Marsyas framework).
 However, I don't think it would be *too* difficult - currently all I
do is calculate the Euclidean distance between two feature vectors
(which contain mostly timbral information, but also some other stuff),
so maybe something could be done like telling the program only to
compare the BPM info when "tempo matching" is chosen.

> I think it may be harder/computationally more expensive to extract than
> timbre information. I know it can be done though. Perhaps you can have a
> look at echonest, which does this, and a *lot* more:
>
> http://developer.echonest.com/
>
> the downside to that is again depending on a 3rd party service, although
> their web API, like last.fm's is free to use for non-commercial use. I'd
> like to do something with it still, perhaps as an additional similarity
> scoring mechanism, but I'm also very interested in automatically mixing
> songs together based on beat detection, which echonest is very good at.
> Anyone who wants to collaborate on such, drop me a note.

Echonest looks pretty awesome, too bad I didn't run across it when I
was working on SoC!  I like that it uses acoustic features rather than
last.fm's system of relying on tags, and relying on a 3rd party
service can't be much worse than relying on a time-consuming analysis
procedure.  Automatically mixing songs is a whole new challenge, but
it would be pretty cool... I guess you would start by choosing a
similar song, then maybe looking at the gaps between beats and
aligning regions of reasonable similarity that are within ~20 seconds
of the ending/beginning of the songs?  Hmm... I really can't commit to
any projects at the moment, but I'd be interested to hear how it comes
along if you do get going on it.

Anyway, fun stuff, and I hope I can help out in the future,

Charlotte