By François Pachet and Pierre Roy.
EURASIP Journal on Audio, Speech, and Music Processing, 2009, article 153017. Publisher record and DOI.
About the paper
The Extractor Discovery System (EDS) builds audio features by composing elementary signal-processing operations. A type system restricts which operations can be combined, so the search can focus on valid expressions for a classification task.
Abstract summary
The paper describes how EDS generates candidate expressions, evaluates them on labelled examples, and uses heuristics and rewriting rules to guide the search. It reports experiments on tasks including music genre classification.
Worked example
One feature in Table 1 applies a Blackman window, splits the signal into frames of 153 samples, and applies a Hann window before measuring spectral spread in each frame. It takes the logarithm of the range of those values. The result is a single number that describes how much spectral spread varies across the excerpt.
Related research
See the Feature Generation project for the research behind EDS, or the Timbre project for related work in music similarity.