DDSC: Dynamic Dual-Signal Curriculum for Data-Efficient Acoustic Scene Classification under Domain Shift

TL;DR

The paper tackles device-induced domain shift in acoustic scene classification under limited labels by introducing Dynamic Dual-Signal Curriculum (DDSC), a training schedule that online-adapts per-example weights using two signals: a domain-invariance score from prototype entropy and a learning-progress score from smoothed loss-change. A time-varying cosine-annealed scheduler fuses these signals to progressively shift emphasis from domain-invariant to device-specific, enabling an easy-to-hard exposure without modifying the model or adding inference overhead. DDSC updates device prototypes online, uses EMA for prototypes, and computes per-example weights via a softmax over a fused score, optimizing a weighted loss $\mathcal{L}_{\mathrm{DDSC}}^{(e)}=\sum_i \pi_i^{(e)} L(x_i,y_i;\theta)$. Evaluated on the DCASE 2024 Task 1 dataset with multiple ASC baselines, DDSC consistently improves cross-device performance, particularly at low-label budgets (e.g., ~4.2% average gain at 5% labels and ~3.9% unseen-device gain), demonstrating data efficiency and practical applicability in domain-general ASC.

Abstract

Acoustic scene classification (ASC) suffers from device-induced domain shift, especially when labels are limited. Prior work focuses on curriculum-based training schedules that structure data presentation by ordering or reweighting training examples from easy-to-hard to facilitate learning; however, existing curricula are static, fixing the ordering or the weights before training and ignoring that example difficulty and marginal utility evolve with the learned representation. To overcome this limitation, we propose the Dynamic Dual-Signal Curriculum (DDSC), a training schedule that adapts the curriculum online by combining two signals computed each epoch: a domain-invariance signal and a learning-progress signal. A time-varying scheduler fuses these signals into per-example weights that prioritize domain-invariant examples in early epochs and progressively emphasize device-specific cases. DDSC is lightweight, architecture-agnostic, and introduces no additional inference overhead. Under the official DCASE 2024 Task~1 protocol, DDSC consistently improves cross-device performance across diverse ASC baselines and label budgets, with the largest gains on unseen-device splits.

Figures (2)

• Figure 1: Device-induced domain shift in ASC. For the same scene, recordings captured with different devices exhibit distinct spectrograms due to device characteristics. Consequently, a model trained on devices A and B may misclassify inputs from an unseen device C, highlighting the challenge of unseen-device generalization.
• Figure 2: Overview of the DDSC. For each epoch, the backbone produces normalized features; the domain-invariance and learning-progress signals are computed and fused by a time-varying scheduler to produce per-example weights used in the weighted objective.