Abstract
In this paper, we investigate DCTNet for audio signal classification. Its output feature is related to Cohen's class of time-frequency distributions. We introduce the use of adaptive DCTNet (A-DCTNet) for audio signals feature extraction. The A-DCTNet applies the idea of constant-Q transform, with its center frequencies of filterbanks geometrically spaced. The A-DCTNet is adaptive to different acoustic scales, and it can better capture low frequency acoustic information that is sensitive to human audio perception than features such as Mel-frequency spectral coefficients (MFSC). We use features extracted by the A-DCTNet as input for classifiers. Experimental results show that the A-DCTNet and Recurrent Neural Networks (RNN) achieve state-of-the-art performance in bird song classification rate, and improve artist identification accuracy in music data. They demonstrate A-DCTNet's applicability to signal processing problems.
Introduction
Effective feature representation is critical for audio classification tasks, such as identifying bird species or music artists. Traditional features like Mel-Frequency Cepstral Coefficients (MFCCs) are widely used due to their alignment with human auditory perception but often suffer from noise sensitivity. Conversely, deep Convolutional Neural Networks (CNNs) can learn powerful features but require massive datasets, extensive tuning, and significant computational resources. Furthermore, CNN filters are often difficult to interpret.
This research introduces the Adaptive DCTNet (A-DCTNet), a hybrid approach that utilizes pre-specified filters based on the Discrete Cosine Transform (DCT). By adapting these filters to the logarithmic nature of acoustic perception, A-DCTNet provides a computationally efficient and mathematically transparent alternative to traditional deep learning models.
Methodology: Adaptive DCTNet
The A-DCTNet architecture builds upon the standard DCTNet by incorporating geometric spacing and multi-layer processing to extract robust features.
Architecture and Geometric Spacing
Unlike standard DCTNet, which uses linearly spaced frequencies, A-DCTNet employs geometrically spaced center frequencies, inspired by the constant-Q transform. This design provides higher resolution for low-frequency components, which are essential for distinguishing timbre and pitch in both human and animal vocalizations.
The process involves:
- Multilayer DCT: The model performs short-time DCT in multiple layers using prefix cosine functions as filters.
- Theoretical Grounding: The authors demonstrate that the output of a two-layer DCTNet belongs to Cohen’s class of time-frequency distributions. This property ensures shift-covariance, making the features suitable for building invariant representations through pooling.
- RNN Integration: To capture the temporal dynamics of audio signals, the extracted features are fed into a Recurrent Neural Network (RNN).
Experimental Results
The model was evaluated on the Artist20 music database (1,413 tracks) and a Bird Song dataset (200 tracks).
Feature Comparison
A-DCTNet features were compared against traditional Log-MFSC features. As shown in the figure below, A-DCTNet provides a more structured representation of the acoustic signal across different scales.
Classification Performance
A-DCTNet consistently outperformed traditional features and the scattering transform:
- Bird Song Classification: The A-DCTNet + RNN combination achieved 98.65% accuracy, surpassing the scattering transform (94.64%) and MFSC (95.26%).
- Artist Identification: On the Artist20 dataset, the model reached 79.11% accuracy, a significant improvement over the baseline MFCC+HMM model (54.49%) and the scattering transform (71.24%).
The results indicate that while A-DCTNet improves performance even with simple linear SVMs, the integration with RNNs is crucial for capturing the sequential information inherent in audio data.
Conclusion
A-DCTNet offers a powerful framework for audio signal processing by combining the interpretability of pre-specified filters with the adaptive power of geometric scaling. It achieves state-of-the-art results in specialized audio tasks while remaining computationally efficient. Future work may explore the impact of deeper architectures and the model's robustness against varying environmental noise levels.
Summary updated