Authors: Vagner Figueira¹
¹University of São Paulo
Background: Hearing comfort and speech intelligibility are highly sensitive to environmental acoustic conditions, and individuals with hearing loss often prefer different amplification settings depending on context [1]. Traditional audiogram-based hearing aid prescriptions may not fully address individual needs across diverse listening environments [2]. Recent surveys on acoustic scene classification (ASC) highlight the potential of AI-driven sound recognition to support adaptive hearing aid technologies by identifying environmental context in real time [3]. AI offers a promising solution by dynamically adjusting signal processing parameters based on the detected scene, enhancing speech clarity, noise suppression, and overall listening comfort—particularly in complex or rapidly changing scenarios [3]. To address this, we present an AI-based acoustic scene recognition system aimed at improving environmental awareness and enabling context-sensitive calibration for hearing-impaired users.
Method: TAU Urban Acoustic Scenes 2020 [2] dataset is used to train a lightweight convolutional neural network (CNN) capable of recognizing environmental scenes. Mel-frequency cepstral coefficients (MFCCs) and their delta features were extracted from the audio recordings as input features. The CNN was developed using Python, Librosa, and Keras, and designed to efficiently classify recordings into three categories: indoor, outdoor, and transportation.
Results: The proposed system achieved an overall classification accuracy of 88% and demonstrated good generalization across different cities and conditions included in the dataset. The model’s lightweight architecture enables its integration into portable devices without compromising performance.
Conclusion: By combining acoustic scene recognition with real-time calibration, our approach has the potential to significantly improve hearing aid functionality and user experience. We are currently developing a cloud-based server infrastructure to process the data in real time and interface with the Tiresias [4] device, enabling adaptive and context-aware hearing aid adjustments in everyday environments.
References
[1] Almufarrij, I., Dillon, H., & Munro, K. J. (2023). Do we need audiogram-based prescriptions? A systematic review. Int. J. of Audiology, 62(6), 500–511.
[2] Bae, J., Park, J., & Kim, H. (2023). Acoustic scene classification: A comprehensive survey. Expert Systems with Applications.
[3] Heittola, T., Mesaros, A., & Virtanen, T. (2020). Acoustic scene classification in DCASE 2020 challenge. Proc. of DCASE2020 Workshop.
[4] Carlet, J.V.C. et al. (2025). Tiresias – An Open-Source Hearing Aid Development Board. AES 158th Convention, Warsaw.
