Audio Oracle: A New Algorithm for Fast Learning of Audio Structures

Shlomo Dubnov (1) , Gerard Assayag (2) , Arshia Cont (3)

1 UC San Diego - University of California [San Diego]
2 STMS - Sciences et Technologies de la Musique et du Son
3 IRCAM - Institut de Recherche et Coordination Acoustique/Musique

In this paper we present a new method for indexing of audio data in terms of repeating sub-clips of variable length that we call audio factors. The new structure allows fast retrieval and recombination of sub-clips in a manner that assures continuity between splice points. The resulting structure accomplishes effectively a new method for texture synthesis, where the amount of innovation is controlled by one of the synthesis parameters. In the paper we present the new structure and describe the algorithms for efficiently computing the different indexing links. Examples of texture synthesis are provided in the paper.

Mots clés: Computer music

Domaines: Son [cs.SD]; Machine Learning [stat.ML]

Shlomo Dubnov, Gerard Assayag, Arshia Cont. Audio Oracle: A New Algorithm for Fast Learning of Audio Structures. Proceedings of International Computer Music Conference (ICMC), 2007, Copenhagen, Denmark. ⟨hal-00839072⟩