Section: New Results
Information-Geometric Approach to Real-time Audio Change Detection
Participants : Arnaud Dessein, Arshia Cont.
We developed a generic framework for real-time change detection of audio signals using methods of information geometry. The present method is limited to generative models of audio signals based on generic exponential distribution families. The proposed system detects changes by controlling the information rate of the signal as they arrive in time. The method also addresses shortcomings of traditional approaches based on cumulative sums which assume known parameters before change. This is achieved by calculating exact generalized likelihood ratio test statistics with complete estimation of unknown parameters in respective hypothesis  . The interpretation of this framework within a dually flat geometry of exponential families provide tractable algorithms for online use. Results are presented for speech segmentation into different speakers and polyphonic music segmentation.