Deep Recurrent Neural Network-Based Autoencoders for Acoustic Novelty Detection.

Area of Science:

• Signal Processing
• Machine Learning
• Acoustics

Background:

• Traditional acoustic novelty detection relies on probabilistic models.
• Recent advancements explore (pseudo-)generative models using recurrent neural networks (RNNs).
• Autoencoders, specifically Long-Short Term Memory (LSTM) denoising autoencoders, predict spectral features for novelty detection via reconstruction error.

Purpose of the Study:

• To address the lack of comprehensive evaluations for RNN-based autoencoders in acoustic novelty detection.
• To compare novel RNN approaches against existing state-of-the-art methods.
• To provide in-depth insights into the performance of autoencoder models for audio event recognition.

Main Methods:

• Utilizing Long-Short Term Memory recurrent denoising autoencoders.
• Predicting auditory spectral features of short-term audio frames.
• Employing reconstruction error as the activation signal for novelty detection.
• Conducting extensive evaluations on three distinct databases: A3Novelty, PASCAL CHiME, and PROMETHEUS.

Main Results:

• RNN-based autoencoders demonstrate superior performance compared to statistical approaches.
• An absolute improvement of up to 16.4% in average F-measure was achieved across the evaluated databases.
• The study provides a consistent and broad evaluation of autoencoder effectiveness in acoustic novelty detection.

Conclusions:

• Recurrent neural network autoencoders represent a significant advancement in acoustic novelty detection.
• These deep learning models offer a more effective alternative to traditional statistical methods.
• The findings highlight the potential of autoencoders for robust audio event recognition and novelty identification.