Contribution
Utilizing the Nonlinear Dynamics of A MEMS Sensor for Acoustic Scene Recognition
* Presenting author
Abstract:
The NeuroSensEar project, supported by the Carl Zeiss Foundation, aims to enhance hearing aid performance and user acceptance by employing Micro Electro-Mechanical System (MEMS) sensors for real-time acoustic scene analysis. The approach involves feeding audio data directly as sound into the sensor, which applies non-linear transformations based on the dynamics of the Euler-Bernoulli beam equation. This equation exhibits a Hopf bifurcation, enabling steady-state and self-resonating behaviors that introduce the necessary non-linearity. The sensor outputs voltage time-series, which are sampled to generate a frequency spectrum, forming the state matrix of the reservoir. Ridge regression, an efficient and analytically solvable method, is then employed to classify or detect the given audio input with high accuracy. Derived from recurrent neural networks, reservoir computing leverages the physical sensor’s natural non-linear processing to minimize power consumption by only training the output layer. This study utilizes the TAU Urban Acoustic Scenes 2022 Mobile development dataset, a widely recognized benchmark for evaluating low-complexity Acoustic Scene Classification models. The goal is to create a low-power hearing aid with an array of MEMS sensors sensitive across a wide range of frequencies, achieving state-of-the-art performance in sound classification and voice detection, and adapting to different environments to improve user experience.