Re: Rhythmbox Predictive Playback - Weekly Report

["Charlotte Curtis" <c f curtis gmail com>]

Good morning everyone,

Last week (week four) I pretty much settled upon a feature extraction
algorithm that I am happy with.  It only uses the middle 30 seconds of
each song which could obviously result in inaccuracies, but doing so
reduces my computation speed significantly - to about 0.5 seconds per
song (in mp3 format) on a fast computer.

I am going to move on to my actual GNOME part of this project: making
the Rhythmbox plugin to call the feature extractor and create
automatic playlists.  I had thought to spend some more time with the
feature extractor to make it write to a database rather than a single
feature file per song, but I think it would be a good idea (as someone
suggested on my blog) to get the plugin working first, and then go
back and make things pretty/optimized.

On Mon, Jun 16, 2008 at 9:36 AM, Charlotte Curtis <c f curtis gmail com> wrote:
> Week Three was quite productive for me.  I implemented a feature
> extraction algorithm using Marsyas that was very successful at last
> years' music information retrieval competition.  I did switch from
> Python to C++, as I saw a significant speed increase, even when
> calling executables from Python code.  I had a minor setback on Friday
> when I managed to overwrite my source code, but it turned out to be
> not a big deal to re-write, as I had learned how the system works.
>
> This week I will work on optimizing the code for speed.  At present, I
> am extracting features for the entire song (most researchers use a
> 30-second segment), and on my Q6600 the process takes about 4.4
> seconds per song.  My old P4 was about 2.2 times slower (based on a
> test that was timed prior to old computer death and repeated on the
> new computer), so I cannot expect that people will want to spend 10
> seconds or more per song extracting features.  My goal is 2 seconds
> (on a moderate processor, tested on my 1.1 Ghz Pentium M), and if I
> can get less than that I will be happy.
>
> That's about it for this week,
>
> Charlotte
>
> On Mon, Jun 9, 2008 at 9:16 AM, Charlotte Curtis <c f curtis gmail com> wrote:
>> I like how people are doing a single thread for all their reports, so
>> I'm going to start doing the same.  Wish I'd thought of it last
>> week...
>>
>> My week began fairly well - I had given myself a deadline of Friday to
>> choose an audio processing library, but after building sample
>> applications and messing around with a few, I decided on Tuesday that
>> I really liked Marsyas.  Not only is Marsyas a library specifically
>> aimed at audio analysis applications, it also has some python
>> bindings, which means I might be able to do everything within a python
>> plugin, provided it's fast enough.  At the very least, I can build and
>> test algorithms in python, and if I need a speed increase, re-create
>> it in C++.
>>
>> Coding became a little bit more difficult on Wednesday morning when my
>> CPU overheated and refused to fire up again.  Now, I could have gone
>> and found a used P4 for sale somewhere, but then I would have had to
>> wait for shipping and all... and besides, it seemed like fate - I
>> bought the computer in September 2003 when I began my undergrad, and
>> it died just over a month after I graduated.  In any case, I live
>> right down the street from (IMO) the best computer shop in Canada
>> (canadacomputers.com!), so I spec'd out a new system, whipped out my
>> shiny new google mastercard, and spent the day on Thursday building,
>> installing, and setting up a new computer.
>>
>> On Friday I managed to get some more work done on my project (four
>> times faster!) - I learned a little about how to use SWIG, and I added
>> bindings to some datatypes of Marsyas that were missing.  The Marsyas
>> devs have been very helpful thus far and I'm glad that I can
>> contribute a little back to them as I work on my project.
>>
>> Cheers,
>>
>> Charlotte
>>
>