Abstract
Implicit Neural Representations (INRs) have seen success in visual domains, but their application to audio—termed Neural Amplitude Fields (NeAF)—remains challenging due to the high-frequency and locally periodic nature of sound. This paper establishes a comprehensive benchmark of Coordinate-MLPs for audio, revealing significant sensitivity to hyperparameters and positional encoding. To address these issues, we introduce Fourier-ASR, a framework based on Kolmogorov-Arnold Networks (KAN) that incorporates Fourier basis functions. Fourier-ASR demonstrates superior robustness, interpretability, and performance, achieving Signal-to-Noise Ratio (SNR) gains of over 12 dB on standard music and speech datasets.
Introduction
Implicit Neural Representations (INRs) represent data as continuous functions mapping coordinates to values. While effective for 3D shapes and images, mapping a timestamp to an audio amplitude is difficult. Audio signals are characterized by high-frequency oscillations and non-linearity. Furthermore, human hearing follows the Weber-Fechner law, making small reconstruction errors highly perceptible. Existing sinusoidal activation models like SIREN require extensive manual tuning. This work provides a systematic study of architectures for audio and proposes a more robust alternative.
Benchmarking Coordinate-MLPs
The authors conducted an extensive evaluation of 48 network configurations, combining various positional encodings with 16 activation functions. This benchmark highlighted two critical weaknesses in traditional MLPs:
- Parameter Sensitivity: Performance is highly dependent on hyperparameter selection. For instance, small changes in the variance scale of positional encodings can lead to dramatic drops in SNR.
- Spectral Bias: Standard MLPs struggle to learn high-frequency details. Positional encoding is not merely beneficial but essential for most activation functions (e.g., ReLU) to capture audio signals effectively.
The Fourier-ASR Framework
To overcome the limitations of MLPs, the authors propose Fourier-ASR, built upon Kolmogorov-Arnold Networks (KAN). Unlike MLPs, which use fixed activations on nodes and learnable weights on edges, KANs place learnable univariate functions on the edges.
Fourier-ASR replaces the standard B-splines used in KANs with Fourier basis functions. This design is naturally aligned with the local periodicity of audio signals, effectively decomposing complex sounds into constituent waves.
To facilitate training, the framework employs a Frequency-adaptive Learning Strategy (FaLS). This includes an "inverted frequency pyramid" initialization that helps the network converge on high-frequency components without losing the low-frequency signal foundation.
Experimental Results
The framework was evaluated on the GTZAN music dataset and the CSTR VCTK speech corpus.
On the GTZAN dataset, Fourier-ASR improved SNR by 12.28 dB over baseline models. On the VCTK speech dataset, the gains reached up to 14.04 dB. Beyond raw performance, Fourier-ASR proved to be significantly more robust than MLPs. While a perfectly tuned MLP (e.g., using NeFF+Sine) can achieve high SNR, it requires "expert-level" hyperparameter hunting. In contrast, Fourier-ASR provides high-quality reconstruction across different datasets with minimal tuning.
Conclusion
This research provides the first comprehensive benchmark for Neural Amplitude Fields and introduces a novel, interpretable framework for audio representation. Fourier-ASR addresses the inherent spectral bias and parameter sensitivity of MLPs by leveraging the periodic properties of Fourier series within a KAN architecture. Future work will explore the application of NeAF to high-fidelity audio compression and downstream tasks such as audio super-resolution and generative synthesis.
Summary updated