Sample-based music composition often involves the task of manually searching appropriate samples from existing audio. Audio mosaicing can be regarded as a way to automatize this process by specifying the desired audio attributes, so that sound snippets that match these attributes are concatenated in a synthesis engine. These attributes are typically derived from a target audio sequence, which might limit the musical control of the user. In our approach, we replace the target audio sequence by a symbolic sequence constructed with pre-defined sound object categories. These sound objects are extracted by means of automatic classification techniques. Three steps are involved in the sound object extraction process: supervised training, automatic classification and user-assisted selection. Two sound object categories are considered: percussive and noisy. We present an analysis/synthesis framework, where the user explores first a song collection using symbolic concepts to create a set of sound objects. Then, the selected sound objects are used in a performance environment based on a loop-sequencer paradigm.

Comunicació presentada a la 6th Sound and Music Computing Conference, celebrada els dies 23 a 25 de juliol de 2009 a Porto, Portugal.

This work is partially funded by Yamaha Corp., Japan and the EU IST project Salero FP6-027122.