Authors: Soumen Sinha¹; Clara Yaiche²
¹TU Delft; ²Absolute Audio Labs
Background: Hearing aid performance depends strongly on the surrounding acoustic environment, where factors such as noise, reverberation, and competing sound sources directly affect speech intelligibility and listening effort. Traditional acoustic scene classification (ASC) systems rely on computationally intensive deep learning models and semantic environment labels (e.g., “street” or “airport”), which are not well aligned with perceptual needs in hearing aid applications and are difficult to deploy on resource-constrained devices. To address these limitations, we propose a hearing-aid–oriented ASC framework that defines scene categories based on acoustic conditions impacting speech perception, enabling more meaningful and adaptive device behavior.
Method: We introduce EdgeConvGRU, a lightweight hybrid architecture combining convolutional neural networks (CNNs) for spectral feature extraction and gated recurrent units (GRUs) for temporal modeling. Audio signals are converted into log-Mel spectrograms to capture perceptually relevant features, and the model processes short audio segments for real-time inference. The architecture is designed with only 34K parameters to ensure low computational complexity, making it suitable for embedded deployment in hearing aids.
Results: The proposed model performed well on benchmark acoustic scene classification tasks, achieving results comparable to larger state-of-the-art models while maintaining a significantly smaller footprint. Despite having only 34K parameters, EdgeConvGRU demonstrates strong classification performance and remains efficient after quantization, highlighting its suitability for resource-constrained deployment.
Conclusion: EdgeConvGRU demonstrates that efficient, perceptually driven acoustic scene classification is feasible for real-time deployment on resource-constrained hearing aids.
