Soutenances

Résumé : Close you eyes, clap your hands. Can you hear the shape of the room? Is the floor made of tiles or carpet? This thesis synthesizes a research journey that attemps to tackle these intriguing questions as engineering problems. Namely, given microphone recordings of a sound source in a room, what can be said geometrically and acoustically not about the source, but about the room? Formalizing this puzzle gives rise to a rich and fascinating network of inverse problems, at the crossroad of computer science, mathematics and physics, most of which remain open to date.

Beside sheer scientific curiosity, making progress on these questions could benefit a number of applications, from simplifying and refining the acoustic diagnosis of rooms, to making audio augmented reality more immersive, to enhancing spatial audio reproduction, to improving the processing of indoor audio signals for teleconferencing, smart devices and hearing aids.

We will present a series of algorithmic contributions to this field, leveraging tools from signal processing, optimization and machine learning, using both models primarily driven by physics and models primarily driven by simulated data. Along the way, some progress will be made in framing the feasibility and well-posedness of these problems, in developing forward and inverse models that strike a balance between complexity and realism, and in understanding the mechanisms that underpin the generalizability of such models to real-world data. Some promising experimental results will be reported on estimating the dimensions, volume, surface area, reflectors' absorption and reverberation time of a room from either impulse response measurements or audio recordings, with or without the aid of geometrical knowledge. We will also investigate the potential benefit of knowing and exploiting such quantities in tasks beyond room acoustics, such as localizing and separating sound sources, and offer directions for future research in the field.