Week 3: Adding timbral similarity

This week, I incorporated the timbral modeling plugin from Queen Mary and created some additional classes to deal with timbral similarity calculations.

I had to modify the Queen Mary SimilarityPlugin to output a vector<float> containing the Gaussian timbral model and beat spectrum. The vector's first three elements give the length of the mean, variance, and beat spectrum vectors respectively, followed by the concatenation of those three vectors.

It turns out that the symmetrised KL divergence doesn't have a specific upper bound, so it's not suitable on its own for combining with other metrics. However, the related Jensen-Shannon divergence is bounded by [0,1], and might be a better fit if it can be efficiently calculated.

As an aside, it seems there is some opportunity for refactoring among the Analyser classes, since AnalyserBeats, AnalyserKey, and AnalyserTimbre all have similar tasks related to serializing protobuf messages to the database and retrieving them. Today I ran into a problem where the TimbreModel was not actually serializing as I had thought, because I bound the data to the query in TrackDAO::bindTrackToLibraryInsert but not in updateTrack or addTracksPrepare (which required almost identical code each time). If these commands were factored out into a separate function, they could easily be applied to queries in all three places.

For next week, I'll investigate the possibilities for such a refactoring, populate the new preference panes with options for ranking the different similarity functions, implement the J-S divergence, and add in the track energy measure as another source of data.