Distance Algorithm?

[evan...@yahoo.com]

I imagine that this is covered somewhere, but I can't seem to find a reference for the algorithm used for distance when a user calls a similarity search. In other words, when sounds are returned as similar with distance 0.45, where does that number come from?

I saw a reference in the 2009 Martinez article to a normalized Manhattan distance calculated from features from the MPG7 standard, bark bands and MFCC. But I didn't see anything saying that this was indeed the algorithm employed to calculate the distance between two sounds from a similarity query to the API.

Is that the case?

If anyone has a better reference on that, I would greatly appreciate a pointer.

Thanks.

[frederic font corbera]

Hello evanxmerz,

you're right that the algorithm is not specified anywhere. 

We're using a standard euclidean distance in a feature vector space containing all lowlevel descriptors listed here: http://www.freesound.org/docs/api/analysis_docs.html#lowlevel-descriptors. Feature vectors are normalized.

We prepared the system to incorporate alternative distance measures in the future, so they could be used as "presets". However, for the moment the only preset available is that one.

However, using the "target" parameter of the content-based search resource (http://www.freesound.org/docs/api/resources.html#sound-content-based-search-resource), you can somehow specify your custom similarity metrics by indicating the features you're interested in. Let me explain better. With this resource you indicate a target value for a list of descriptors, and then the system will return the most similar sounds ONLY taking into account the listed descriptors (also using Euclidean distance). You can not weight the relevance of the different features, but you can restrict which features to consider. Nevertheless, this resource is still in testing, and does not seem to work well for long lists of descriptors (specially when using the filter parameter too).

Thank you for your interest! If you have any suggestions don't hesitate to post them.

frederic

2013/3/7 <evan...@yahoo.com>

--
 
---
You received this message because you are subscribed to the Google Groups "Freesound api" group.
To unsubscribe from this group and stop receiving emails from it, send an email to freesound-ap...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.