From CNN to Reservoir Computing: A New Perspective on Acoustic Scene Classification

He Y, Hosseini AMA, Abeßer J, Jaurigue L, Raake A, Lüdge K. From CNN to Reservoir Computing: A New Perspective on Acoustic Scene Classification. In: 2025 AES International Conference on Artificial Intelligence and Machine Learning for Audio. Audio Engineering Society; 2025. Paper 21. Available from: https://aes.org/publications/elibrary-page/?id=23010

@inproceedings{He2025_23010,
  author = {He, Yuxuan and Hosseini, Alireza Molla Ali and Abeßer, Jakob and Jaurigue, Lina and Raake, Alexander and Lüdge, Kathy},
  title = {{From CNN to Reservoir Computing: A New Perspective on Acoustic Scene Classification}},
  booktitle = {2025 AES International Conference on Artificial Intelligence and Machine Learning for Audio},
  note = {Paper 21},
  year = {2025},
  month = sep,
  publisher = {Audio Engineering Society},
  url = {https://aes.org/publications/elibrary-page/?id=23010}
}

TY  - CPAPER
TI  - From CNN to Reservoir Computing: A New Perspective on Acoustic Scene Classification
AU  - He, Yuxuan
AU  - Hosseini, Alireza Molla Ali
AU  - Abeßer, Jakob
AU  - Jaurigue, Lina
AU  - Raake, Alexander
AU  - Lüdge, Kathy
T2  - 2025 AES International Conference on Artificial Intelligence and Machine Learning for Audio
M1  - Paper 21
PY  - 2025
DA  - 2025/09/02
UR  - https://aes.org/publications/elibrary-page/?id=23010
PB  - Audio Engineering Society
LA  - en
AB  - Acoustic Scene Classification (ASC) is a fundamental task in machine listening, aiming to categorize the acoustic environment in which an audio recording was made. Although Convolutional Neural Networks (CNNs) are still the dominant approach in ASC today, increasing attention has been given to optimizing model efficiency, particularly for model deployment in resource-limited computational platforms such as hearing aids. In this study, we explore Reservoir Computing (RC) as an alternative lightweight approach for ASC and systematically compare its performance against multiple CNN architectures, including ResNet-34, MobileNet-V2, EfficientNet-B2, and TF-SepNet-20. Our results indicate that CNN models consistently outperform RC in classification accuracy, even when trained without data augmentation. However, RC exhibits significantly lower computational cost, requires less training time, and achieves the fastest inference speed among all models. This trade-off between accuracy and efficiency suggests that model selection should be guided by application requirements: CNNs are preferable when accuracy is paramount, while RC provides an alternative for low-power real-time applications such as hearing aids.
ER  -