Introduction
Acoustic Scene Classification (ASC)—the task of identifying environments like parks, buses, or malls from audio—is traditionally performed by converting signals into 2D spectrograms. While effective, this approach often discards phase information crucial for spatial hearing and leads to over-parameterized 2D-CNNs that are too large for low-power edge devices. This paper demonstrates that lightweight 1D-CNNs trained directly on raw binaural waveforms can achieve state-of-the-art performance while remaining small enough for extreme hardware constraints.
Methodology
The proposed pipeline moves from raw audio to a highly compressed deployment model through specialized architectures and rigorous optimization.
Architectures
The authors evaluated two 1D-CNN variants: a VGG-like stack of convolutional blocks and a Residual Network (ResNet) architecture adapted for 1D signals. Unlike standard models that mono-mix audio, these networks utilize both left and right channels to exploit inter-aural phase differences.
The ResNet variant incorporates residual blocks to facilitate deeper feature extraction while maintaining stability during training, which is particularly useful for capturing complex temporal patterns in raw audio.
Data Augmentation and Compression
To prevent overfitting on limited data, the researchers applied five augmentation techniques: temporal masking, frequency filtering, additive Gaussian noise, Mixup, and CutMix. For deployment, the models were optimized using:
- Structured Pruning: Iteratively removing the least important convolutional kernels based on L1 norm.
- Quantization: Converting weights from 32-bit to 16-bit precision, halving the memory footprint with negligible accuracy loss.
Experimental Setup
The study utilized the TAU Urban Acoustic Scenes 2020 3class dataset, consisting of binaural recordings from various European cities categorized into indoor, outdoor, and transportation classes. Audio was downsampled to 18 kHz to reduce computational load while preserving relevant signal energy.
Results and Discussion
The experimental results show that 1D-CNNs can outperform traditional 2D-CNN baselines while being significantly more efficient.
| Model | Accuracy | Size (KB) | Notes |
|---|---|---|---|
| Baseline | 87.2% | 220.0 | Standard spectrogram-based 2D-CNN |
| Model A | 87.6% | 23.8 | 1D-CNN; 20x smaller than the limit |
| Model B | 90.9% | 58.4 | Larger 1D-CNN; high accuracy/size ratio |
| Model D | 91.2% | 466.6 | Pruned ResNet; best overall accuracy |
Key Insights
- Phase Matters: Binaural input provided a +5.9% accuracy boost for raw waveform models, compared to only +1.6% for spectrogram models. This suggests 1D-CNNs effectively learn to use inter-aural phase differences that spectrograms typically ignore.
- End-to-End Efficiency: Model A achieved baseline accuracy with a footprint 20 times smaller than the competition's 500 KB limit, proving that "end-to-end" audio learning is viable for low-power devices.
- Pruning Resilience: ResNet models maintained high accuracy even after two-thirds of their parameters were pruned, indicating significant natural over-parameterization for 3-class problems.
Conclusion
This research demonstrates that raw-waveform 1D-CNNs are a powerful alternative to traditional spectrogram-based methods for low-complexity ASC. By leveraging binaural phase information and aggressive compression, the authors created models as small as 24 KB that outperform standard baselines. Future work will explore scaling these architectures to larger datasets and analyzing the specific features learned by the 1D convolutional filters.
Summary updated