Section: New Results

Real-time Multi-object Detection for Music Signals

Participants : Philippe Cuvillier [Master 2 ATIAM] , Arshia Cont.

Multiple-object detection and tracking has been widely used in applications such as missile tracking and radar and has given birth to several formalisms such as Random Finite Sets [33] . Such formalisms can be seen as extensions to existing probabilistic inference mechanisms with explicit birth and death stochastic mechanisms for multiple source tracking.

In this work we aim at studying such formalisms in the case of real-time music signal processing. The idea is to track multiple sources (instruments, audio flows) from one source of observation. This approach can be beneficial to two main applications in real-time music listening:

• Extension of existing audio-to-score [2] or audio-to-audio alignment [7] mechanisms (currently based on one source) to multiple objects can address the following short-comings of existing approaches: explicit consideration for asynchrony of parallel sources; robustness to uncertainties on one or more voices.

• Studying the classical Partial Tracking applications in audio processing within the RFS context can lead to better results in low-level sinusoidal partial tracking of sounds.

Early studies of such formalisms are exposed in [25] . Concrete applications will be exposed in 2013.