Deep learning for sound source localization and orchestral music source separation

Speaker: David Diaz-Guerra

4 Marzo 2026 | 14:00
DEIB, Sala Seminari "A. Alario" (Ed. 21)

Contatti:  Prof.  Mirco Pezzoli

Sommario

On March 4^th, 2026, at 2:00 pm the seminar on "Deep learning for sound source localization and orchestral music source separation" will take place in DEIB “Alessandra Alario” Seminar Room (Building 21).

In the last decade, deep learning has revolutionized most fields of audio signal processing, and sound source localization (SSL) and sound source separation (SSS) are not exceptions. However, relevant challenges remain open. Despite their better localization accuracy, SSL models still struggle to work in highly reverberant environments, and there is a need for the development of smaller models suitable for operating in real time. Following the ideas of the Geometric Deep Learning, understanding the symmetries of the problem can help to design more robust and efficient models. In the field of SSS, new deep-learning models have been proposed for speech and for music separation, though the latests usually focus on popular music with a reduced number of sources. SSS for classical music remains a challenging task due to the lack of data and the high number of sources and spectral overlap.

Biografia

David Diaz-Guerra studied telecommunications engineering at the University of Zaragoza (Spain), where he obtained his Ph.D. degree in March 2023 with his thesis “A Geometric Deep Learning Approach to Sound Source Localization and Tracking”. He is now a postdoctoral researcher at the Audio Research Group of Tampere University (Finland), where he worked on sound source separation within the European Research and Innovation (RIA) project REPERTORIUM, and he is now conducting his own research project on spatial audio funded by the Tampere Institute for Advanced Study (IAS). He is also a member of the organizing team of the Task 3 (sound source localization and detection) of the DCASE Challeng since 2023.