Adaptive Knowledge Distillation using a Device-Aware Teacher for Low-Complexity Acoustic Scene Classification

2025-09-11

AI Summary

This work addresses acoustic scene classification on low-complexity devices under strict computational constraints, aiming to enhance model robustness across both seen and unseen devices. To this end, we propose a device-aware feature alignment loss, a dual-teacher knowledge distillation framework—integrating PaSST and a device-specialized teacher—and lightweight test-time fine-tuning guided by device labels. The student model, CP-MobileNet, ensures low inference latency while enabling efficient knowledge transfer. Experiments demonstrate that our approach achieves 57.93% accuracy on the development set, significantly outperforming the official baseline. Notably, it exhibits substantially improved generalization to unseen devices, validating the effectiveness of device-aware distillation and test-time adaptive fine-tuning.

Technology Category

• Machine Learning: Hardware-aware ML
• Computer Vision: Learning & Optimization for CV
• Search and Optimization: Learning to Search

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Search and Retrieval-Augmented AI: Personalized, context-aware and across-device search

Abstract

In this technical report, we describe our submission for Task 1, Low-Complexity Device-Robust Acoustic Scene Classification, of the DCASE 2025 Challenge. Our work tackles the dual challenges of strict complexity constraints and robust generalization to both seen and unseen devices, while also leveraging the new rule allowing the use of device labels at test time. Our proposed system is based on a knowledge distillation framework where an efficient CP-MobileNet student learns from a compact, specialized two-teacher ensemble. This ensemble combines a baseline PaSST teacher, trained with standard cross-entropy, and a 'generalization expert' teacher. This expert is trained using our novel Device-Aware Feature Alignment (DAFA) loss, adapted from prior work, which explicitly structures the feature space for device robustness. To capitalize on the availability of test-time device labels, the distilled student model then undergoes a final device-specific fine-tuning stage. Our proposed system achieves a final accuracy of 57.93% on the development set, demonstrating a significant improvement over the official baseline, particularly on unseen devices.

Problem

• Low-complexity acoustic scene classification under strict constraints
• Robust generalization to both seen and unseen recording devices
• Leveraging test-time device labels through knowledge distillation framework

Innovation

• Knowledge distillation with device-aware teacher ensemble
• Device-Aware Feature Alignment loss for robustness
• Device-specific fine-tuning using test-time labels