Low-Complexity Acoustic Scene Classification with Device Information in the DCASE 2025 Challenge

TL;DR

This work defines a low-complexity acoustic scene classification task for DCASE 2025 in which the recording device ID is available at inference, enabling device-aware specialization under tight resource constraints. A two-stage baseline is proposed: first learn a general model from a 25% data subset, then adapt to known devices via per-device fine-tuning, using mel-spectrogram features and a receptive-field-regularized, factorized CNN; Freq-MixStyle augmentation addresses device mismatch. The TAU Urban Acoustic Scenes 2022 Mobile dataset provides the evaluation backbone, with external data allowed to support transfer learning. On the development-test split, the general model achieves $50.72\%$ accuracy, which improves to $51.89\%$ when using device-specific adaptation, illustrating the value of device-aware modeling for low-resource ASC. The setup emphasizes practical deployment on embedded hardware and identifies transfer learning as a promising avenue for further gains.

Abstract

This paper presents the Low-Complexity Acoustic Scene Classification with Device Information Task of the DCASE 2025 Challenge and its baseline system. Continuing the focus on low-complexity models, data efficiency, and device mismatch from previous editions (2022--2024), this year's task introduces a key change: recording device information is now provided at inference time. This enables the development of device-specific models that leverage device characteristics -- reflecting real-world deployment scenarios in which a model is designed with awareness of the underlying hardware. The training set matches the 25% subset used in the corresponding DCASE 2024 challenge, with no restrictions on external data use, highlighting transfer learning as a central topic. The baseline achieves 50.72% accuracy on this ten-class problem with a device-general model, improving to 51.89% when using the available device information.

Figures (1)

• Figure 1: Overview of Low-Complexity Acoustic Scene Classification with Device Information. At inference time, models must operate under low-complexity constraints and handle both known (seen during training) and unknown (unseen during training) recording devices, with the device ID provided. The baseline follows a two-stage training process: first, learning a general model, then adapting it to device-specific characteristics to enhance performance on known devices.