DNN-Based Sound Event Detection with Exemplar-Based Approach for Noise Reduction

In Kyu Choi, Kisoo Kwon, Soo Hyun Bae, Nam Soo KimDCASE 2016 · OpenReview

Abstract

Sound Event Detection (SED) in real-world environments faces significant hurdles, including overlapping sound events (polyphony), limited annotated datasets, and unpredictable background noise. This paper proposes a robust SED system using a Deep Neural Network (DNN) combined with a novel exemplar-based noise reduction technique. By leveraging aggressive data augmentation to synthesize polyphonic scenes and dynamically extracting noise profiles from test audio, the system achieves high performance on the DCASE 2016 benchmark, particularly in noisy office environments.

Introduction

Sound Event Detection (SED) involves identifying the onset and offset of specific acoustic events within an audio stream. While monophonic detection is relatively straightforward, real-world scenarios are characterized by polyphony—where multiple events occur simultaneously—and varying levels of background noise. Traditional models often struggle with these complexities due to data sparsity and the inability to generalize to unseen noise types. This research addresses these issues through a multi-faceted approach involving synthetic data generation, in-situ noise estimation, and adaptive classification thresholds.

System Architecture

The proposed system is structured into two primary stages: a training stage and a test stage. During training, the focus is on building a robust classifier using augmented data. In the test stage, the system processes real-world audio through a pipeline of feature extraction, noise reduction, and DNN-based classification.

Figure 1: Overview of the proposed SED system architecture, detailing the distinct training and test stages.

Exemplar-Based Noise Reduction

A core innovation of this work is the exemplar-based noise reduction mechanism. Unlike static methods that rely on pre-defined noise models, this approach estimates noise "in-situ" from the test recording itself. The system identifies frames with the lowest 30% energy, assuming these segments represent the ambient background noise. These frames are stored as "noise exemplars."

Figure 2: The exemplar-based noise reduction workflow, showing how noise is sampled and subtracted to enhance features.

For every incoming audio frame, the system finds the best-matching noise exemplar and subtracts its mel-band energy from the frame's features. This dynamic subtraction effectively cleans the signal of unpredictable background noise before it reaches the classifier.

Data Augmentation and DNN Classification

To combat data sparsity, the authors employ aggressive data augmentation. Isolated sound events are randomly time-stretched ($\pm 10\%$) and their power is varied ($50\% - 200\%$). These modified events are then overlapped to create synthetic polyphonic scenes, which are further mixed with various noise types (white, pink, and blue).

The resulting acoustic features, incorporating an 11-frame context window, are fed into a DNN with three hidden layers. The output layer utilizes sigmoid activation functions for multi-label classification, enabling the simultaneous detection of multiple sound classes. Furthermore, an adaptive thresholding mechanism is used during post-processing. By raising the detection threshold for sound classes that share acoustic similarities with the estimated background noise, the system significantly reduces false-positive rates.

Experimental Results

The system was evaluated using the IEEE DCASE 2016 Challenge Task 2 dataset, which focuses on 11 office-related sound classes. The results demonstrate the effectiveness of the proposed pipeline:

  • Development Set: The system achieved a segment-based F-score of 0.9261 and an Error Rate (ER) of 0.1379. The more stringent event-based F-score was 0.7939.
  • Evaluation Set: On unseen data, the system maintained strong performance with a segment-based F-score of 0.787 and an ER of 0.3660. The event-based F-score reached 0.671.

These results highlight the system's ability to maintain high precision even when faced with novel acoustic conditions and overlapping events.

Conclusion

This paper demonstrates that combining in-situ noise estimation with synthetic data augmentation provides a powerful framework for polyphonic SED. The exemplar-based noise reduction technique proves particularly effective at handling unpredictable environments. Future work may focus on narrowing the gap between segment-based and event-based performance by refining the detection of event onsets and offsets in highly dense acoustic scenes.

Summary updated

BibTeX
@inproceedings{choi2016dnn,
  author = {In Kyu Choi and Kisoo Kwon and Soo Hyun Bae and Nam Soo Kim},
  title = {{DNN-Based Sound Event Detection with Exemplar-Based Approach for Noise Reduction}},
  booktitle = {DCASE},
  year = {2016},
  url = {https://openreview.net/forum?id=MqExuA6Ltm},
}
Figure