Abstract
Sound event location is a critical aspect of two-dimensional direction-of-arrival (2D-DOA) estimation, predicting azimuth and elevation angles in 3D Cartesian coordinates for active sound events using multi-label regression. Challenges with conventional methods like the multi-signal classification (MUSIC) algorithm and baseline convolution recurrent neural network (BCRNN) include decreased precision and high computational demands, particularly in low signal-to-noise ratio (SNR) environments (SNR < -5 dB). Our work introduces an innovative solution, the effective residual self-attention recurrent neural network (ESRNN). ESRNN addresses distortion problems in low SNR conditions caused by the MUSIC algorithm, also enhancing 2D-DOA prediction accuracy in various SNR-reverberation scenarios. We propose two filter structures, ESRNN-L and ESRNN-G, tailored for SNRs above 0 dB and below -5 dB, respectively. Evaluating on TAU Spatial Sound Events 2019 datasets with synthetic SNRs from -10 dB to 30 dB, our experiments demonstrate ESRNN-L achieves a 21% lower 2D-DOA error than BCRNN at SNRs below -5 dB. Additionally, ESRNN-G exhibits a 15% lower error with a 10% parameter reduction when SNRs exceed 0 dB. When compared with other principal attention methods through ablation study, it also showcases the model's efficiency and robustness.
Introduction
Locating the source of a sound—technically known as Direction-of-Arrival (DOA) estimation—is a fundamental task for technologies ranging from smart-home assistants and autonomous vehicles to industrial acoustic monitoring. In ideal conditions, this is a straightforward geometric problem. However, real-world environments introduce significant background noise and "multipath" interference where sound waves bounce off surfaces.
Traditional algorithms like MUSIC (Multiple Signal Classification) are mathematically elegant but brittle, often failing when the Signal-to-Noise Ratio (SNR) drops below -5 dB. While modern deep learning approaches like Convolutional Recurrent Neural Networks (BCRNNs) are more robust, they often struggle with precision and high computational costs in extreme noise. The ESRNN architecture was developed to bridge this gap, maintaining high accuracy even when noise levels exceed the signal strength.
Methodology
The Effective Residual Self-Attention Recurrent Neural Network (ESRNN) is built upon three primary technical pillars designed to enhance signal clarity and focus:
1. Soft Thresholding (Residual Shrink Block)
Inspired by signal denoising techniques, the network utilizes a "soft thresholding" function. The model learns an adaptive threshold for each feature channel; features falling below this threshold (likely noise) are suppressed, while significant signal features are preserved. This allows the network to dynamically "zero out" noise-related interference in real-time.
2. Residual Self-Attention
Standard attention mechanisms help models focus on important sequence segments. By adding residual (skip) connections, ESRNN ensures that the model does not lose original signal information as the network depth increases, effectively mitigating the vanishing gradient problem.
3. Specialized Model Variants
To optimize for different environments, the researchers proposed two variants:
- ESRNN-L (Local): Optimized for high-noise environments (SNR < -5 dB), it filters noise layer-by-layer to preserve fine-grained local speech features.
- ESRNN-G (Global): Designed for clearer environments (SNR > 0 dB), it utilizes a global context vector to capture broader patterns with higher computational efficiency.
Experimental Setup and Results
The study utilized the TAU Spatial Sound Events 2019 dataset, testing both Ambisonic (FOA) and Microphone Array (MIC) formats. Researchers synthesized additional noise levels ranging from a clean 30 dB to a punishing -10 dB to test the limits of the architecture.
Key findings include:
- Superior Accuracy in Noise: In high-noise scenarios (SNR < -5 dB), ESRNN-L achieved a 21% lower 2D-DOA error compared to the BCRNN baseline.
- Efficiency Gains: In low-noise scenarios (SNR > 0 dB), ESRNN-G exhibited a 15% lower error while reducing the total parameter count by 10% (down to 6.43MB).
- Robustness: Ablation studies comparing ESRNN to other popular attention modules (such as SE, CBAM, and ECA) showed that ESRNN consistently provided the highest accuracy gains.
Conclusion
ESRNN demonstrates that noise suppression can be treated as a learnable parameter within a deep learning framework, allowing for high-precision sound location in environments where classical signal processing fails. Future work will focus on expanding the model to handle multi-tasking—simultaneously detecting the type of sound event while locating it—and testing against non-Gaussian impulse noise.
Summary updated