DNN Transfer Learning based Non-linear Feature Extraction for Acoustic Event Classification

Seongkyu Mun, Minkyu Shin, Suwon Shon, Wooil Kim, David K. Han, Hanseok Ko

Abstract

Recent acoustic event classification research has focused on training suitable filters to represent acoustic events. However, due to limited availability of target event databases and linearity of conventional filters, there is still room for improving performance. By exploiting the non-linear modeling of deep neural networks (DNNs) and their ability to learn beyond pre-trained environments, this letter proposes a DNN-based feature extraction scheme for the classification of acoustic events. The effectiveness and robustness to noise of the proposed method are demonstrated using a database of indoor surveillance environments.

Introduction

Acoustic Event Classification (AEC)—the ability for a system to autonomously recognize sounds like breaking glass, footsteps, or sirens—is critical for surveillance and smart home applications. However, AEC faces two major hurdles: data scarcity and linearity constraints. Unlike speech recognition or computer vision, AEC lacks massive, high-quality labeled datasets, making it difficult to train complex models from scratch. Furthermore, traditional acoustic features (like MFCCs) or learned filters (like NMF) are largely based on linear combinations, which often fail to capture the complex, non-linear relationships between frequency bands in real-world sounds.

This paper proposes a novel feature extraction scheme that leverages Deep Neural Networks (DNNs) and transfer learning to "borrow" knowledge from large, diverse audio sets and apply it to specific, data-poor target tasks.

Proposed Method: DNN-based Non-linear Feature Extraction

The core innovation is treating a DNN as a sophisticated, learned non-linear filter bank rather than just a classifier. The process involves three main stages: source domain pre-training, transfer and adaptation, and feature extraction.

Comparison of conventional filter-training and the proposed DNN-based approach.

Initially, a DNN is trained on a massive "source" dataset to learn general acoustic representations. This pre-trained network captures broad patterns across many sound classes. Subsequently, the output layer of this network is removed and replaced with new adaptation layers tailored to the specific target task, such as indoor surveillance. The network is then fine-tuned on the target domain data.

Detailed architecture of the transfer learning process, showing source and target domain tasks with specific node counts.

Once trained, the final activation functions and output layer are stripped away. The remaining hidden layers act as the "DNN filter." The raw values from the last hidden layer (a bottleneck approach) are extracted as the new acoustic features. These features are then processed via standard techniques like Discrete Cosine Transform (DCT) for dimensionality reduction and fed into a final classifier, such as an SVM.

Experimental Evaluation

The study utilized a two-stage data approach:

  • Source Domain: A merged dataset of 93 classes from CLEAR-OL, RWCP, UrbanSound, and ESC-50 to provide a broad background of acoustic knowledge.
  • Target Domain: An indoor surveillance database consisting of 15 event types (e.g., crying child, gunshot, explosion).
  • Noise Robustness: Background noise from the ETSI database was added at varying Signal-to-Noise Ratios (5, 10, and 15 dB) to test real-world performance.

Key Results

The proposed method demonstrated a clear advantage over both traditional and other deep-learning-based features:

  • Superior Accuracy: The proposed "DNN filter + SVM + DCT" setup achieved an average accuracy of 95.8% across clean and noisy conditions.
  • Outperforming Baselines: It significantly beat standard MFCCs (88.5%), NMF (91.9%), and K-SVD (92.5%).
  • Noise Robustness: In high-noise environments (5 dB SNR), the proposed method maintained 92.5% accuracy in living room noise, whereas MFCC performance dropped to 79.7%.
  • Transfer Learning Gain: Comparing a model trained from scratch to the transferred model showed a jump from 93.0% to 95.8%, proving that parameters learned from general environmental sounds provide a powerful initialization for specific surveillance events.

Conclusion

By utilizing the non-linear modeling of DNNs and the efficiency of transfer learning, this research provides a robust framework for acoustic event classification even with limited target data. Future work will investigate how this scheme holds up across other diverse acoustic domains and the impact of different input feature configurations on performance.

Summary updated

Abstract

Recent acoustic event classification research has focused on training suitable filters to represent acoustic events. However, due to limited availability of target event databases and linearity of conventional filters, there is still room for improving performance. By exploiting the non-linear modeling of deep neural networks (DNNs) and their ability to learn beyond pre-trained environments, this letter proposes a DNN-based feature extraction scheme for the classification of acoustic events. The effectiveness and robustness to noise of the proposed method are demonstrated using a database of indoor surveillance environments.
BibTeX
@article{mun2017dnn,
  author = {Seongkyu Mun and Minkyu Shin and Suwon Shon and Wooil Kim and David K. Han and Hanseok Ko},
  title = {{DNN Transfer Learning based Non-linear Feature Extraction for Acoustic Event Classification}},
  journal = {IEICE Trans. Information and Systems},
  year = {2017},
  url = {https://openreview.net/forum?id=-xJnPskED0},
}
Figure