We invite you to join us to our monthly Artificial Intelligence Meeting (AIM) at Observatoire astronomique de Strasbourg.
—> Monday October 21st, 11:00-12:00, in the Amphitheatre of the Observatoire astronomique de Strasbourg (and Zoom, link below).
Paul Magron (INRIA, Université de Lorraine )will talk about Demixing Sounds with AI: towards deep phase recovery
Abstract:
Sound demixing consists in automatically extracting the constitutive components (the sources) that add-up to form an observed audio recording (the mixture). This task is of paramount importance in applications such as speech enhancement for hearing aids, augmented music mixing, or audio scene analysis.
In this talk, I will first introduce and illustrate sound demixing throughout applications, and present a general pipeline for addressing this problem using modern AI approaches. These usually transform the sound into a time-frequency representation that is made up of a spectrogram and a phase, and process the spectrogram via deep neural networks.
Then, I will focus on the specific sub-problem of phase recovery, whose goal is to alleviate the above issue of processing spectrogram-only quantities. I will propose a model of the phase structure via signal analysis, and design optimization-based iterative phase recovery algorithms.
Finally, I will present my current research on the topic of deep phase recovery, where phase models are now deeply learned instead of hand-crafted, and where optimization algorithms are unfolded into neural networks for end-to-end training instead of fixed for post-processing. Indeed, combining deep learning with traditional signal processing tools is a promising research direction for maintaining high performance while improving audio processing systems' robustness and interpretability.
Biography:
Since 2021, Paul Magron is a tenured research scientist (Chargé de Recherche) with INRIA centre at Université de Lorraine. He previously completed his PhD at Télécom Paris (2016), and worked as a postdoctoral researched at Tampere University (Finland, 2017-2019), and IRIT (Toulouse, 2019-2021). His research interests include audio signal processing and machine learning, with applications to sound demixing, speech enhancement, and audio restoration.
———————————————
Zoom link:
Join Zoom Meeting
Meeting ID: 964 4923 5172
Passcode: 5q04wV
———————————————
Cheers,
Paolo, Olivier, Julien
This event is co-funded by the IRMIA++ Young Researchers Budget.
