| Résumé |
Audio segmentation is an essential problem in many audio signal processing tasks which tries to segment an audio signal into homogeneous chunks, or segments. Most current approaches rely on a change-point detection phase for finding segment boundaries, followed by a similarity matching phase which identifies similar segments. In this thesis, we focus instead on joint segmentation and clustering algorithms which solve both tasks simultaneously, through the use of unsupervised learning techniques in sequential models. Hidden Markov and semi-Markov models are a natural choice for this modeling task, and we present their use in the context of audio segmentation. We then explore the use of online learning techniques in sequential models and their application to real-time audio segmentation tasks. We present an existing online EM algorithm for hidden Markov models and extend it to hidden semi-Markov models by introducing a different parameterization of semi-Markov chains. Finally, we develop new online learning algorithms for sequential models based on incremental optimization of surrogate functions. |
