Introduction
Acoustic Scene Classification (ASC) is a fundamental task in audio signal processing, aimed at identifying the environment in which a sound was recorded—such as a busy street, a quiet park, or a bustling restaurant. While significant progress has been made in device-invariant ASC, a new frontier has emerged: geographical domain shift. This paper introduces the IEEE ICME 2024 Grand Challenge, which addresses the performance degradation encountered when models trained on data from one region (e.g., Europe) are deployed in another (e.g., China).
The Challenge of Geographical Domain Shift
Geographical domain shift occurs because the "acoustic signature" of a scene is deeply tied to local culture, language, architecture, and daily habits. A model trained exclusively on European acoustic data may achieve high accuracy locally but fail significantly when tested on Chinese recordings. Preliminary findings show that a model trained on European data achieved a mere 14% accuracy in Chinese environments, highlighting a massive "culture gap" in audio data.
Beyond the shift itself, the field faces a "data scarcity" bottleneck. Labeling audio data is labor-intensive and expensive. To combat this, the challenge promotes the use of semi-supervised learning to leverage vast amounts of unlabeled audio data.
Semi-Supervised Methodology
The proposed baseline system utilizes a multi-stage semi-supervised framework designed to adapt a general model to a specific local domain using limited labeled data.
The process follows four key steps:
- Pre-training: An initial SE-Trans model is trained on the TAU UAS 2020 dataset.
- Fine-tuning: The model is refined using a small set of labeled data from the target Chinese cities.
- Pseudo-labeling: The model generates "pseudo-labels" for a large pool of unlabeled Chinese audio.
- Final Training: The model undergoes a final training phase using both the original labeled data and the high-confidence pseudo-labeled examples to improve robustness.
The architecture, SE-Trans, integrates Squeeze-and-Excitation (SE) blocks to emphasize critical features and a Transformer encoder to model long-term temporal dependencies in the audio signal.
The CAS 2023 Dataset
The challenge is built upon the Chinese Acoustic Scene (CAS) 2023 dataset, which provides a distinct cultural profile compared to previous European-centric datasets.
- Scale: Over 130 hours of audio across 10 scene categories from 22 Chinese cities.
- Development Set: Contains 4.8 hours of labeled data and 19.3 hours of unlabeled data from 8 cities.
- Evaluation Set: Includes 3 hours of data from 12 cities, 5 of which are "unseen" during training to rigorously test domain adaptation.
Results and Scene-Specific Insights
The baseline system achieved a macro-average accuracy of 59% on the evaluation set. However, performance varied significantly across different environments. Scenes with distinct, rhythmic, or consistent sounds, such as Metro (90%) and Bar (87%), were classified with high precision. Conversely, complex outdoor environments like Public Squares (29%) proved difficult, as they often lack a unique acoustic signature and blend into a generic mix of voices and distant traffic.
Conclusion
The ICME 2024 Grand Challenge sets a new benchmark for robust ASC. While the baseline provides a solid foundation, the significant performance gap in "unseen" cities and confusing scenes like public squares remains a hurdle. The research community is invited to develop more sophisticated semi-supervised algorithms and domain adaptation techniques to bridge the geographical divide in acoustic scene understanding.
Summary updated