At SONY CSL Music we have pioneered the field of audio feature generation: building, through evolutionary techniques, signal processing functions that can be used as efficient features for music classification.
We introduced the notion of analytical features: audio features consisting in mathematical compositions of elementary operators. Our tool, called EDS (Extractor Discovery System) is able to design automatically a fully-fledged audio extractor (an executable), given only a test database and corresponding perceptive tests. Specific and efficient audio features are found using genetic programming, and aggregated together using machine-learning algorithms, for which optimized parameters are found. These extractors are readily useable in the MusicBrowser application.
The technology of EDS is now transferred to Sony Corporation for many applications in the field of signal processing.
This video illustrates how EDS can learn to discriminate between guitar and piano sounds, finding appriopriate signal processing functions (features) and then applying it to music corpora.