Research Article
A Supervised Domain Adaptive Method for Multi Device Acoustic Scene Classification
@INPROCEEDINGS{10.1007/978-3-031-60347-1_2, author={Zhanqi Liu and Mei Wang and Liyan Luo and Zhenghong Liu and Guan Wang}, title={A Supervised Domain Adaptive Method for Multi Device Acoustic Scene Classification}, proceedings={Mobile Multimedia Communications. 16th EAI International Conference, MobiMedia 2023, Guilin, China, July 22-24, 2023, Proceedings}, proceedings_a={MOBIMEDIA}, year={2024}, month={10}, keywords={Acoustic Scene Classification Supervised Domain Adaptive Domain generalization Frequency band standardization}, doi={10.1007/978-3-031-60347-1_2} }- Zhanqi Liu
Mei Wang
Liyan Luo
Zhenghong Liu
Guan Wang
Year: 2024
A Supervised Domain Adaptive Method for Multi Device Acoustic Scene Classification
MOBIMEDIA
Springer
DOI: 10.1007/978-3-031-60347-1_2
Abstract
Acoustic scene classification faces performance degradation due to device mismatch when different pickup devices are used in the training and testing phases. To solve the device mismatch problem and improve the performance of the acoustic scene classification system on unseen pickup devices, we propose a joint frequency band standardization and supervised domain adaptation algorithm, which can effectively extract the domain invariant features of the acoustic scene signal to generalize the model to the unseen The algorithm can effectively extract the domain invariant features of the acoustic scene signal and generalize the model to the unseen distribution to solve the performance degradation of the model on the unseen pickups. The frequency band standardization is first used to linearly correct the extracted Log-Mel features, and then combined with supervised domain adaptation to reduce the difference between the source and target domains and correct the nonlinear differences between different kinds of pickup devices. Experimental results on the DCASE Challenge 2020 Task 1A dataset show an overall improvement of 13.9% compared to the baseline model, 6.2% on seen devices, and 20% on unseen device. The algorithm can better extract the domain invariant features of the model, has better classification performance, and enables the model to generalize to unseen device.
