This week, I worked on implementing the Last.fm tag fetcher, to add an additional data source for determining song similarity. Since the official liblastfm is covered under GPL v3 and Mixxx uses GPL v2, I switched from using the liblastfm library to connecting to the REST API using Qt's built-in functions.
In order to fetch tags, you can currently use the "Fetch tags..." item in the context menu -- it will later be something that happens during the analysis, if it's enabled in the Selector preferences (not yet implemented). In addition, I added a menu item to "Get follow-up tracks" which will set the currently selected track as a seed for the selector and switch over to Selector view.
The fetcher obtains results from Last.fm fine, but I want to make sure that if any tracks have no tags or could not be found (as was the case for several tracks in my test) that we store that information, so that Last.fm will not be repeatedly queried for them (if they have been updated recently).
Max notes that upon opening Mixxx with a newly created database, the Selector view is empty. I can confirm this is happening, but I'm still not sure why; this is something important to look into further.
As for my solution to using the Hellinger distance for timbral similarity, Max has suggested that I implement a test to compare the success of different distance functions using this data. For this purpose, we can use the ISMIR '04 dataset which is freely available online to assess how well the metric groups together similar tracks.
Next week I'll work on putting such a test in place, as well as implementing the tag similarity functionality and isolating a "get best follow-up" function so it can be called from e.g. the auto-dj.
No comments:
Post a Comment