Bag-of-Features Acoustic Event Detection for Sensor Networks

Julian Kürby, Rene Grzeszick, Axel Plinge, Gernot A. FinkDCASE 2016 · OpenReview

Introduction

Acoustic Event Detection (AED) is the task of identifying specific sounds—such as glass breaking, footsteps, or phone rings—within complex, noisy environments. While single-microphone systems have seen significant advancement, they often struggle with "cluttered" acoustic scenes or sounds occurring at a distance. Sensor networks, consisting of distributed microphone arrays, provide a promising solution by offering multiple spatial "viewpoints" of the same event.

The primary challenge in utilizing sensor networks lies in effective data fusion: how to intelligently combine inputs from dozens of microphones. Traditional heuristic methods, such as majority voting or simple averaging, often fail to account for the varying confidence levels of individual sensors. Furthermore, complex models can become overly specialized to a specific room layout, losing efficacy when sound sources move to unobserved locations. This paper proposes a robust multi-channel AED framework based on the "Bag-of-Features" (BoF) approach, utilizing classifier stacking and channel sorting to achieve superior, location-agnostic performance.

Methodology: Multi-Channel Bag-of-Features

The proposed system extends the standard Bag-of-Features pipeline to a multi-sensor environment. The process is divided into local processing for each microphone and a centralized fusion stage.

Overview of the multi-channel Bag-of-Features architecture.

Local Processing

For each microphone channel, the system performs the following steps:

  1. Feature Extraction: 27-dimensional feature vectors are extracted, comprising Gammatone Frequency Cepstral Coefficients (GFCCs) and loudness filters.
  2. Quantization: A Gaussian Mixture Model (GMM) is employed for "soft assignment" of features to "acoustic words." This results in a histogram-based BoF representation for each time window.
  3. Classification: A Maximum Likelihood classifier estimates the posterior probability of each event class for that specific microphone.

Learned Fusion via Stacking

The core innovation lies in the fusion stage. Instead of applying fixed heuristic rules, the system employs classifier stacking. The posterior probabilities from all microphones are fed into a secondary Random Forest meta-classifier. This model learns to recognize patterns in the collective sensor outputs, effectively identifying which combinations of microphone signals indicate a true event.

To ensure the system remains robust to changes in sound source location, the authors introduce channel sorting. Before being fed into the stacking classifier, microphone channels are sorted based on their highest posterior probability. This makes the model "location-agnostic" by teaching it to prioritize the most confident sensors rather than specific physical microphone IDs.

Experimental Evaluation

The approach was validated using two datasets featuring 32 synchronized microphones:

  • ITC-Irst Database: A benchmark with 16 acoustic event classes recorded using T-shaped microphone arrays.
  • TU Dortmund Smart Room: A challenging dataset with 19 categories, specifically designed to test generalization across different room locations ("separate" vs. "mixed" training/testing sets).

Results and Discussion

The experimental results demonstrate that learned fusion significantly outperforms traditional methods.

Comparison of F-scores for various fusion methods.

Key Findings

  • Channel-Specific vs. Global Models: Training individual BoF models for each microphone location (channel-specific) yields better results than a single "global" model. This highlights the importance of capturing the unique local acoustic characteristics of each sensor's position.
  • Superiority of Stacking: As shown in the performance comparison, the stacking method consistently achieves the highest F-scores compared to heuristic fusion rules like majority voting or taking the maximum probability. On the ITC-Irst dataset, stacking reached a frame-based F-score of 84.2%.
  • Generalization through Sorting: In scenarios where training and testing sound locations were strictly separated, the channel sorting technique reduced classification error from 10.0% to 7.1%. This confirms that sorting allows the stacking classifier to generalize to new spatial configurations.
  • Computational Efficiency: The system is highly efficient, running at approximately 20% of real-time on a standard single-core i7 CPU, making it suitable for practical, low-latency deployment.

Conclusion

By combining the Bag-of-Features framework with learned classifier stacking and channel sorting, this research provides a robust solution for multi-channel acoustic event detection. The system effectively leverages the spatial diversity of sensor networks while maintaining high performance across varying sound source locations. Future work will address the challenges of asynchronous data and varying sensor counts in ad-hoc networks, where perfect frame-level synchronization may not be guaranteed.

Summary updated

BibTeX
@inproceedings{kurby2016bag,
  author = {Julian Kürby and Rene Grzeszick and Axel Plinge and Gernot A. Fink},
  title = {{Bag-of-Features Acoustic Event Detection for Sensor Networks}},
  booktitle = {DCASE},
  year = {2016},
  url = {https://openreview.net/forum?id=dhEKgQs4an},
}
Figure