Abstract
Sound Source Localization (SSL) involves estimating the Direction of Arrival (DOA) of sound sources. Since the DOA estimation output space is continuous, regression might be more suitable for DOA, offering higher precision. However, in practice, classification often outperforms regression, exhibiting greater robustness to interference. Conversely, classification's drawback is inherent quantization error. Within the classification paradigm, the DOA output space is discretized into intervals, each treated as a class. These classes show strong inter-class correlations, being inherently ordered, with higher similarity as intervals grow closer. To address this, we propose an Unbiased Label Distribution (ULD) to eliminate quantization error in training targets. Furthermore, we tailor two loss functions for the soft label family: Negative Log Absolute Error (NLAE) and Mean Squared Error without activation (MSE(wo)). Finally, we introduce Weighted Adjacent Decoding (WAD) to overcome quantization error during model prediction decoding. Experimental results demonstrate our approach surpasses classification quantization limits, achieving state-of-the-art performance.
Introduction
In Sound Source Localization (SSL), researchers typically choose between two paths to estimate the Direction of Arrival (DOA): regression and classification. Regression treats the angle as a continuous number, which is precise but often struggles with noise and interference. Classification divides the 360° circle into discrete "bins" (e.g., every 5°). While classification is much more robust in reverberant environments, it suffers from inherent quantization error. For a bin width $l$, the expected error floor is $l/4$. This work introduces a framework to eliminate this "quantization tax" while preserving the robustness of classification.
Methodology
The proposed framework bridges the gap between classification and regression by exploiting the spatial correlation between adjacent bins. Unlike standard classification tasks where categories are independent, SSL classes are inherently ordered.
Unbiased Label Distribution (ULD)
Instead of a "one-hot" label where only one bin is active, ULD uses "soft" labels. If a sound is at 12.4° and bins are at 10° and 15°, ULD distributes probability between them based on the exact decimal distance. This creates an injective mapping, allowing the model to represent angles more precisely than the bin resolution.
Tailored Loss Functions and Decoding
Standard functions like Cross-Entropy (CE) do not handle soft labels optimally. We introduce Negative Log Absolute Error (NLAE) and Mean Squared Error without activation (MSE(wo)) to provide better gradients for training. During inference, Weighted Adjacent Decoding (WAD) calculates a weighted average of the peak bin and its neighbors to interpolate the exact position, allowing the model to point to locations between predefined classes.
Key Insights and Results
A fundamental insight is that SSL bins are spatially correlated. As shown in Figure 2, ULD provides a continuous relationship with spatial distance (measured via Wasserstein distance) that discrete one-hot encoding lacks.
Performance Highlights
- Breaking the Quantization Limit: In anechoic tests with 5° bins (theoretical limit 1.25°), the proposed method achieved a Mean Absolute Error (MAE) of just 0.061°—nearly 20 times better than the theoretical floor.
- Real-World Robustness: On real-world office recordings, the ULD + MSE(wo) + WAD-3 combination reduced MAE from 3.168° to 2.925° compared to standard classification.
- Efficiency: In simulated scenarios, the strategy reduced MAE by up to 45.22% over industry-standard setups (One-hot + CE + Top-1).
Limitations and Conclusion
The approach is most effective when the model can accurately identify the correct neighborhood. In extremely noisy environments, "classification error" (picking the wrong bin entirely) can outweigh quantization error. Furthermore, making bins too small (e.g., 1°) can lead to overfitting. By combining the robustness of classification with the precision of regression, this framework effectively breaks the theoretical quantization floor of SSL systems.
Code and supplementary materials: https://github.com/linfeng-feng/ULD
Summary updated