A Hypernetwork-Based Approach to KAN Representation of Audio Signals

Patryk Marszalek, Maciej Rut, Piotr Kawa, Piotr Syga

Introduction

Implicit Neural Representations (INRs) have emerged as a powerful paradigm for data modeling, representing signals as continuous functions parameterized by neural network weights. While INRs have seen significant success in 3D computer vision, audio representation remains a formidable challenge. Sound signals are characterized by high-frequency components and dense temporal information, often leading to "spectral bias" in standard Multi-Layer Perceptrons (MLPs), where the network prioritizes learning low-frequency structures at the expense of high-fidelity reconstruction.

Methodology: Audio-KAN and FewSound

To address these limitations, this research explores the application of Kolmogorov-Arnold Networks (KANs) to audio signals. Unlike traditional MLPs that utilize fixed activation functions at nodes, KANs employ learnable activation functions—specifically B-splines—on the edges. This architectural shift allows the network to more effectively capture the complex, non-linear shapes inherent in audio waveforms.

Comparison between the Fourier Transform and the Kolmogorov-Arnold Network architecture, highlighting the shift from fixed basis functions to learnable splines.

As illustrated above, the KAN approach can be viewed as a deep learning evolution of classical signal processing techniques like the Fourier Transform. While a Fourier Transform decomposes a signal into a sum of fixed sine waves, a KAN learns an optimal set of basis functions (splines) to represent the signal. This allows for a more efficient representation of the "shape" of the signal.

Furthermore, the researchers introduced FewSound, a hypernetwork-based meta-learning framework. Training an INR for every individual audio clip is computationally expensive. FewSound utilizes an encoder to process raw audio and predict weight updates for a base KAN model, enabling high-fidelity "few-shot" reconstruction with a single forward pass.

Experimental Results

The proposed models were evaluated across diverse datasets, including VCTK (speech), GTZAN (music), and UrbanSound8K (environmental noise).

Perceptual Quality

A key finding is that KANs excel in perceptual metrics. While traditional Mean Squared Error (MSE) measures mathematical divergence, metrics like PESQ (Perceptual Evaluation of Speech Quality) better reflect human hearing. KANs achieved a state-of-the-art (SOTA) PESQ score of 3.57 for 1.5s clips, significantly outperforming NeRF and SIREN-based models. This suggests that the learnable splines in KANs are particularly adept at capturing the natural textures and harmonic structures of sound that human ears prioritize.

FewSound Performance

The FewSound hypernetwork demonstrated substantial improvements over the previous state-of-the-art, HyperSound:

  • 33.3% reduction in Mean Squared Error (MSE).
  • 60.87% improvement in Scale-Invariant Signal-to-Noise Ratio (SI-SNR).

Discussion and Limitations

The success of KANs in audio representation stems from their resistance to spectral bias. By utilizing learnable basis functions, they can represent high-frequency details that standard MLPs often "blur" out. However, challenges remain regarding computational complexity, as KANs can be more resource-intensive than simple MLPs as sequence length increases. Additionally, scaling these representations to longer audio sequences (beyond 3 seconds) and ensuring cross-linguistic robustness are areas for future investigation.

Conclusion

This work demonstrates that Kolmogorov-Arnold Networks, combined with hypernetwork-based meta-learning, provide a superior framework for high-fidelity audio representation. By moving beyond fixed activations to learnable splines, KANs offer a more flexible and perceptually accurate method for digital signal processing in the neural era.

Summary updated

BibTeX
@inproceedings{marszalek2025hypernetwork,
  author = {Patryk Marszalek and Maciej Rut and Piotr Kawa and Piotr Syga},
  title = {{A Hypernetwork-Based Approach to KAN Representation of Audio Signals}},
  booktitle = {ACM International Conference on Information and Knowledge Management},
  year = {2025},
  doi = {10.1145/3746252.3761405},
}
Figure