Section: New Results

Online Automatic Structure Discovery of Audio Signals

Participants : Arshia Cont, Vincent Lostanlen [MS Internship] .

Following recent team findings in [12] and the framework introduced in [4] , we pursued the problem of automatic discovery of audio signals using methods of information geometry through a Masters Thesis undertaken by Vincent Lostanlen (MS ATIAM) [34] . This work introduces a novel way of representing and calculating Similarity Matrices for continuous multimedia signals and in real-time. In this approach, the signal is first segmented into homogeneous chunks using the change detection algorithm proposed by the team in [12] , and proposes a method for constituting similarity relations between segments using Bregman Information Geometry and exploiting intersections between information balls.

Compared to traditional approaches to similarity matrix computing, the approach proposed in [34] is strictly on-line (thus suitable for real-time computing) and provides a sparse view of audio structures. We will pursue this project by increasing its robustness and evaluating results on larger databases including other timed-signals such as video.