Sound Scene Understanding Team Seminar (Talk by Dr.Diego DI CARLO).

This is an online seminar. Registration is required.

【Sound Scene Understanding Team】
【Date】2022/Jan/20(Thu) 1630pm-1730pm (JST)

【Speaker】Dr.Diego DI CARLO

Title: "Echo-aware signal processing for audio scene analysis... and pills of Physics-Informed Neural Network"
Abstract: Close your eyes, clap your hands, can you "hear" the shape of the room? This particular question was addressed by a famous work that demonstrates the possibility of recovering the shape of a room from the analysis of acoustic echoes.
This idea was later extended to many significant audio inverse problems, (e.g. sound source localization, speech enhancement, and others), and the term "echo-aware methods" have been proposed.
In this seminar, I will present the main results of my PhD work, where I tried to lay a unified analytical and methodological foundation for addressing two questions: how to estimate and use such acoustic echoes properly?
Such results concern two methods for echoes estimation from stereo recordings (one based on continuous dictionaries, one on neural networks) and the extension of existing methods to their echo-aware forms: the SRP-PHAT localization method, the Multichannel NMF framework for audio source separation, and the MVDR beamformer for speech enhancement. Moreover, I will present a novel dataset specifically designed and annotated for echo-aware applications.
Finally, the neural network presented for echoes estimation can be viewed as a "physics-informed neural network" (PINNs), where physical laws act as prior knowledge to design robust algorithms. In my current research, I am studying PINNs for image super-resolution of turbulent flows, but the fundamental ideas can be applied to audio information processing.