DANN beats CDAN for device-invariant acoustic scene classification
New study shows DANN adapts well across devices, but CDAN struggles with transformers.
A new study from researchers Abhishek Dileep, Shubham Sharma, and Padmanabhan Rajan investigates how domain adaptation techniques can make acoustic scene classification models invariant to different recording devices. The team tested two popular methods—DANN (Domain Adversarial Neural Network) and CDAN (Conditional Domain Adversarial Network)—across two types of feature extractors: convolutional neural networks (CNNs) and transformer-based models. Using the DCASE 2020 dataset, which contains audio from multiple devices, they evaluated how well each method handles domain shifts caused by different microphones and hardware.
The results reveal a clear performance gap: DANN provided effective domain adaptation consistently for both CNN and transformer extractors, making it a robust choice for device-invariant classification. In contrast, CDAN only performed well when paired with CNN-based feature extractors, failing to adapt transformers adequately. This suggests that CDAN's conditional alignment mechanism is more sensitive to the underlying feature representation, especially with transformer architectures that capture global context differently. The 6-page paper, published on arXiv (2607.25887), offers practical guidance for engineers building acoustic scene classification systems that need to work reliably across varied hardware.
- DANN adapts consistently across both CNN and transformer feature extractors for acoustic scene classification.
- CDAN only works well with CNN extractors, failing to generalize to transformer-based models.
- Study uses the DCASE 2020 multi-device dataset to evaluate domain shifts in real-world recording conditions.
Why It Matters
Enables more robust acoustic AI systems that work across devices without retraining on each hardware setup.