Adaptive DCTNet for audio signal classification

Yin Xian, Yunchen Pu, Zhe Gan, Liang Lu, Andrew Thompson

TL;DR

The authors introduce A-DCTNet, an adaptive version of the Discrete Cosine Transform Network designed specifically for audio classification. By using geometrically spaced filters (inspired by the constant-Q transform) and pairing them with Recurrent Neural Networks (RNNs), they achieve state-of-the-art results in bird song classification and music artist identification while maintaining a model that is more interpretable and computationally efficient than traditional deep CNNs.

What Problem Does This Paper Address?

In the world of audio classification—think identifying a bird species from a recording or a musician from a song—the "feature representation" is everything. Traditional features like Mel-Frequency Cepstral Coefficients (MFCCs) are popular because they mimic human hearing, but they are notoriously sensitive to noise.

On the other end of the spectrum, Convolutional Neural Networks (CNNs) are powerful but come with heavy baggage: they require massive amounts of data, expert-level parameter tuning, and significant computational power to "learn" filters that are often difficult to interpret. The authors sought a middle ground: a system that uses "pre-specified" filters (like the scattering transform or PCANet) that are "learned for free" but adapted specifically to the logarithmic nature of human acoustic perception.

Methods

The core innovation is the Adaptive DCTNet (A-DCTNet). Here’s how it works:

  1. Multilayer DCT: Instead of learning filters from scratch, the model uses a prefix cosine function as a filter. It performs a short-time Discrete Cosine Transform (DCT) in multiple layers.
  2. Geometric Spacing: Standard DCTNet uses linearly spaced frequencies. A-DCTNet switches to geometrically spaced center frequencies (a "constant-Q" approach). This provides much higher resolution for low-frequency components, which is where the most critical information for human (and many animal) audio perception resides.
  3. Theoretical Grounding: The authors mathematically prove that the output of a two-layer DCTNet belongs to Cohen’s class of time-frequency distributions. This is significant because it means the features are shift-covariant, making them ideal for building invariant features through pooling.
  4. RNN Integration: To capture the temporal "flow" of audio (how sounds change over time), the features extracted by A-DCTNet are fed into a Recurrent Neural Network (RNN) rather than just a static classifier.

Datasets

The model was evaluated on two distinct types of audio data:

  • Artist20: A music database containing 1,413 tracks from six albums by 20 different artists. The goal is to identify the artist based on 30-second clips.
  • Bird Song Data: A dataset from Duke University’s Nowicki Lab featuring 200 tracks across four different bird species. These clips are short (2-3 seconds) and require capturing fine-grained acoustic details.

Key Insights and Ideas

  • "Learned for Free": Because the filters are based on the DCT, they don't need to be trained via backpropagation. This makes the feature extraction phase incredibly fast and mathematically transparent compared to a CNN.
  • Phase Matters: The authors suggest that A-DCTNet outperforms the popular "Scattering Transform" because it preserves phase information in each layer, whereas scattering transforms typically take absolute values early on, potentially discarding useful data.
  • Human-Centric Engineering: By aligning the filter banks with geometric scales, the model naturally focuses on the frequency ranges that distinguish timbre and pitch, effectively "hard-coding" biological intuition into the architecture.

Key Results

A-DCTNet consistently outperformed traditional features (MFSC, LFSC, ERB scale) and the more complex scattering transform across both datasets:

  • Bird Song Classification: A-DCTNet paired with an RNN achieved a staggering 98.65% accuracy, significantly beating the scattering transform (94.64%) and MFSC (95.26%).
  • Artist Identification (Artist20): The A-DCTNet + RNN combo reached 79.11% accuracy. For context, the baseline MFCC+HMM model only reached 54.49%, and the scattering transform reached 71.24%.
  • SVM vs. RNN: While A-DCTNet improved results even with a simple linear SVM, the jump in performance when using an RNN highlights the importance of capturing sequential information in audio signals.

Limitations and Future Work

While the paper demonstrates strong results, a few areas remain open for exploration:

  • Architecture Depth: The study focuses primarily on a two-layer DCTNet. It is unclear if adding more layers would provide diminishing returns or if the mathematical properties (like the connection to Cohen's class) would hold up or become too complex.
  • Noise Robustness: Although the authors mention that traditional features are sensitive to noise, they did not provide a rigorous stress test of A-DCTNet against varying levels of environmental or white noise in this specific paper.
  • Hyperparameter Sensitivity: The choice of "filters per octave" ($b$) was set to 12 for the first layer and 6 for the second. Future work could investigate if these should be learned or if there is a universal "sweet spot" for different types of audio (e.g., speech vs. industrial noise).

Summary updated

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.
BibTeX
@inproceedings{xian2017adaptive,
  author = {Yin Xian and Yunchen Pu and Zhe Gan and Liang Lu and Andrew Thompson},
  title = {{Adaptive DCTNet for audio signal classification}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2017},
  url = {https://openreview.net/forum?id=Z3iZFGtE9wJ},
}

Loading related content...

Figure