8–12 Sept 2026
Europe/Vienna timezone

Online Estimation of Virtual Sensing Observation Filter Coefficients with Real-World Sounds for Local Active Noise Control

FA2026/559
11 Sept 2026, 09:00
3h
Messehalle (Poster+Exhibition) (Messe Congress Graz)

Messehalle (Poster+Exhibition)

Messe Congress Graz

A09 Machine learning and artificial intelligence in acoustics A09.01/A17.01 Machine learning for array processing

Speaker

Felix Holzmüller (Institute of Electronic Music and Acoustics)

Description

Adaptive local active noise control (ANC) requires accurate estimates of the sound pressure at the point of cancellation. These are often obtained using nearby microphones via virtual sensing methods such as the remote microphone technique (RMT). Obs-TasNet was recently proposed as a neural network–based method for online estimation of RMT observation filter coefficients. In this work, we evaluate Obs-TasNet's performance in synthesized reverberant environments with stochastic as well as nonstationary sounds sampled from the FSD50K dataset. Overall, Obs-TasNet exhibits only minor degradation of estimation error when applied to nonstationary sounds compared to stochastic noise sources. In reverberant environments, performance declines toward higher frequencies approaching the array's aliasing frequency. Nevertheless, Obs-TasNet consistently outperforms an inverse-distance-weighting baseline. These results indicate that Obs-TasNet is robust across source types and can handle acoustic conditions relevant to practical deployment.

Authors

Felix Holzmüller (Institute of Electronic Music and Acoustics) Paul Armin Bereuter (Institute of Electronic Music and Acoustics) Christian Blöcher (Institute of Electronic Music and Acoustics) Alois Sontacchi (University of Music and Performing Arts)

Presentation materials