Seminar from Vincent Lostanlen, researcher at LS2N

Date : 05/09/2025
Time : 10h00
Place : IC2, Boardroom
Speaker : Vincent Lostanlen
 
 

Machine learning for bioacoustics: Challenges and opportunities

 
Abstract

The recent development of inexpensive digital audio sensors, wireless communications, and high-performance computing have opened new horizons in ecology, conservation, and animal behavior. Yet, the adoption of passive acoustic monitoring technologies by practitioners in life sciences is hindered by some pressing challenges: namely, class imbalance, reliability, interpretability, environmental responsibility, and privacy preservation. In this talk, i will review these challenges individually and discuss some “learned lessons”, having myself engaged in bioacoustics research since 2015. Then, i will present some emerging techniques to address these challenges: few-shot learning, edge computing, solar-powered hardware, per-channel energy normalization, and training set synthesis. These techniques are being studied by two PhD students at LS2N: Ms. Yasmine Benhamadi and M. Matthieu Carreau.

Biography

He completed his PhD in 2017 at École normale supérieure, then worked as a postdoctoral researcher at the Cornell Lab of Ornithology and New York University’s Music and Audio Research Lab. Since 2020, he is a scientist at CNRS, the French national center for scientific research, where he co-advises PhD students on signal processing and neural network methods for representation learning in audio, with applications to bioacoustics and digital music. He also serves as computer music designer for artist Florian Hecker, with creations for the Lausanne Underground Film Festival, BBC Radio 3, Editions Mego, Urbanomic, Hyperdub, and Warp Records. Team website: https://audio.ls2n.fr