Sound event localization and classification using WASN in Outdoor Environment

Introduction

This paper presents a novel deep learning-based method for sound event localization and classification (SELC) using wireless acoustic sensor networks (WASN) in outdoor environments. The key aspects are:

• WASN are used for signal acquisition and processing, with multiple microphone arrays to enable comprehensive monitoring of large outdoor areas.

• A Soundmap feature is proposed, which leverages the geometric information of the array to enhance spatial gain and suppress noise interference, enabling effective extraction of spatial information in low signal-to-noise environments.

• Gammatonegram is used to represent the sound signals, as it aligns better with human auditory characteristics and is more effective in outdoor settings.

• A multitask model based on convolutional neural networks and Transformer encoder modules is introduced, which integrates sound event classification and sound source localization tasks.

• The proposed method is evaluated in diverse acoustic environments and shows superior performance compared to state-of-the-art deep learning-based approaches.

PROPOSED METHOD

The text describes a system for simultaneous sound source localization (SSL) and sound event classification (SEC) using a wireless acoustic sensor network (WASN). The key components are:

Soundmap Features: The system uses broadband beamforming to obtain spatial-energy information across the frequency spectrum, creating a "Soundmap" feature that captures the frequency-domain distribution of the target sound signal.

Gammatonegram (GTGram) Features: To handle acoustic environments with noise and interference, the system uses Gammatone filterbanks to generate perceptually-relevant GTGram features.

Neural Network Architecture: The system uses a deep neural network with convolutional and transformer encoder layers to learn the mapping between the Soundmap, GTGram, and array position features to the sound source location and classification.

The paper elaborates on the mathematical formulations for the Soundmap and GTGram feature extraction, as well as the details of the neural network modules. The final model outputs the estimated sound source coordinates for SSL and the sound event categories for SEC.

The loss function combines mean squared error for SSL and binary cross-entropy for SEC, with a weighting parameter to balance the two objectives.

EVALUATION

This section introduces simulated multi-array and multi-channel datasets, as well as evaluation metrics for the proposed method and baseline methods.

The authors use the Pyroomacoustic package to simulate sound signal propagation. The experimental setup includes 8-element circular microphone arrays with a radius of 11 cm and a sampling frequency of 8000 Hz. Three types of sound events are considered: emergency siren, human scream, and gunshot, along with a noise category. The simulated data covers five different area sizes, with various sound pressure levels (SPLs) and signal-to-noise ratios (SNRs) for the target and interfering sources.

For the sound event classification (SEC) task, the authors use precision, recall, F1 score, and false alarm rate (FAR) as evaluation metrics. For the sound source localization (SSL) task, they use root mean square error (RMSE) as the metric. Additionally, a comprehensive metric called the SELC score is designed, which combines the F1 score, 1-FAR, and 1-RMSE/area length.

The authors also provide implementation details for the Soundmap and GTGram features, as well as the array position feature. Finally, they introduce several baseline methods for comparison, including SEC-CNN, SEC-SEC, SSL-PLSE, SSL-FUZZY, and SSL-STFT.

V RESULTS AND DISCUSSION

The paper evaluates the baseline methods and the proposed method from various perspectives. The key findings are:

SEC performance: The proposed method outperforms two baseline methods in sound event classification (SEC), achieving higher F1 scores and lower false alarm rates across different sound classes.

SSL performance: The proposed method also outperforms three baseline methods in sound source localization (SSL), with an average localization error of 7.5 meters, significantly lower than the other methods.

Ablation study: Removing the SEC branch from the proposed network decreases localization accuracy, indicating SEC improves SSL performance.

Comparison of SELC: The proposed method achieves the highest combined score for sound event localization and classification, outperforming two baseline methods that use a cascaded or multitask approach.

Real-world experiment: In a real-world urban park setting, the proposed method maintains superior performance in both SEC and SSL tasks compared to the baselines, demonstrating its practicality.

CONCLUSION

This paper presents a deep learning-based method for sound event localization and classification using a wireless acoustic sensor network (WASN). The approach uses multiple microphone arrays, each with multiple sensors, to sample and process acoustic signals.

The key innovations include a novel feature called Soundmap that represents spatial information across multiple frequency bands, and a network architecture that employs attention mechanisms to learn channel-wise relationships and temporal dependencies in the acoustic features.

Experimental results show the proposed method outperforms baseline approaches. It achieves the highest F1 score and lowest false alarm rate for the sound event classification task, as well as the lowest root-mean-square error for the sound source localization task. Further experiments confirm the mutual benefits of jointly learning these two tasks.

Visualization of localization errors demonstrates the robust sound source localization performance of the proposed method in complex environments. The efficiency of the approach is also validated through real-world experiments.