Query-by-Multiple-Examples: Content-Based Search in Computer-Assisted Sound-Based Musical Composition

We propose a search method, namely Query-by-MultipleExamples, that is able to search, within an audio sample database, for a particular sonic characteristic. The characteristic is learned on-the-fly by means of multiple examples provided by a human user, thus avoiding ambiguities due to manual labelling. We evaluate four variations of the proposed method using ground truth provided by three musicians. It is shown that, for queries based on sonic characteristics, the query modelling process yields more correct results than if several single-example queries were executed in parallel using the same input data.