A Scale-Free Distribution of False Positives for a Large Class of Audio Similarity Measures

By Jean-Julien Aucouturier and François Pachet.

Pattern Recognition 41(1), 272–284 (2008). DOI: 10.1016/j.patcog.2007.04.012.

Read the paper (PDF) · Publisher version (DOI)

About the paper

Why do some songs keep turning up as close matches in a music search, even when they sound unlike the query? This paper follows up on observations from the authors’ earlier work on timbre similarity. In a collection of 15,460 tracks, a few songs appeared in thousands of nearest-neighbor lists. The counts followed an approximately power-law distribution.

Abstract summary

Aucouturier and Pachet tested several bag-of-frames models that summarize the spectral features of a recording. Hub songs appeared across model variants, but changing the algorithm could change which songs became hubs. Their experiments traced the effect to how models combined features across audio frames. In their comparison, polyphonic music exhibited hubness while the environmental recordings they tested did not.

Related work

Hubness in Music Similarity connects this study to the earlier timbre papers and later research on hubness.

Improving Timbre Similarity: How High’s the Sky? reports the earlier experiments that identified recurring false positives.

Music Similarity Measures: What’s the Use? introduces the timbre similarity measure that motivated these experiments.

“The Way It Sounds” describes the timbre models used for music analysis and retrieval.

The Bag-of-Frames Approach to Audio Pattern Recognition compares the effectiveness of bag-of-frames models for recognizing urban soundscapes and polyphonic music.

Back to Publications