Contributions in Audio Modeling and Multimodal Data Analysis via Machine Learning
Résumé
This HDR manuscript summarizes our work concerning the applications of machine learning techniques to solve various problems in audio and multimodal data. First, we will present nonnegative matrix factorization (NMF) modeling of audio spectrograms to address audio source separation problem, both in single-channel and multichannel settings. Second, we will focus on the multiple instance learning (MIL)-based audio-visual representation learning approach, which allows to tackle several tasks such as event/object classification, audio event detection, and visual object localization. Third, we will present contributions in the multimodal multimedia interestingness and memorability, including novel dataset constructions, analysis, and computational models. This summary is based on major contributions published in three journal papers in the IEEE/ACM Transactions on Audio, Speech, and Language Processing, and a paper presented at the International Conference on Computer Vision (ICCV). Finally, we will briefly mention our other works in different applications concerning audio synchronization, audio zoom, audio classification, audio style transfer, speech inpainting, and image inpainting.
Origine : Fichiers produits par l'(les) auteur(s)
Loading...