Representing Sounds as Neural Amplitude Fields: A Benchmark of Coordinate-MLPs and A Fourier Kolmogorov-Arnold Framework
[AAAI 2025]

Linfei Li · Lin Zhang* · Zhong Wang · Fengyi Zhang · Zelin Li · Ying Shen

🌐Project page | 📝Paper (AAAI) | 📝Paper (arXiv)

What is Neural Amplitude Fields?

Environment Configuration

• Configure a Python environment and install related dependencies.

  pip install -r requirements.txt
  

• Download the required dataset from the following websites.

  • CSTR VCTK: https://datashare.ed.ac.uk/handle/10283/3443
  • GTZAN: https://huggingface.co/datasets/marsyas/gtzan
  • Bach: https://github.com/vsitzmann/siren/blob/master/data/gt_bach.wav
  • Counting: https://github.com/vsitzmann/siren/blob/master/data/gt_counting.wav

• Organize the data set according to the following file structure.

  --data
    --demo
      --gt_bach.wav
      --gt_counting.wav
      --gt_blues00000.wav  # from GTZAN dataset blues_00000.wav
    --gtzan
      --genres
        --blues
        ...
    --VCTK
      --wav48_silence_trimmed
        --p231
        ...
  

Benchmark of Coordinate-MLPs in Audio Signal Representations

Overview

Run

• Testing on Bach, Counting, and Blues.

bash scripts/benchmark_MLPs_demo.sh

• Testing on CSTR VCTK dataset.

bash scripts/benchmark_MLPs_vctk.sh

• Testing on GTZAN dataset.

bash scripts/benchmark_MLPs_gtzan.sh

Fourier-ASR: A Fourier Kolmogorov-Arnold Framework

Overview

Run

• Testing on Bach, Counting, and Blues.

bash scripts/benchmark_KANs_demo.sh

• Testing on CSTR VCTK dataset.

bash scripts/benchmark_KANs_vctk.sh

• Testing on GTZAN dataset.

bash scripts/benchmark_KANs_gtzan.sh

Ablation Experiments

Positional encoding is parameter-sensitive

• RFF positional encoding is sensitive to the dimension parameter $L$.

bash scripts/benchmark_FFN_L.sh

• RFF positional encoding is sensitive to the variance parameter $\sigma$.

bash scripts/benchmark_FFN_sigma.sh

• NeFF positional encoding is sensitive to the dimension parameter $L$.

bash scripts/benchmark_NeRF_L.sh

Activation functions are parameter-sensitive

• Gaussian-type activation functions are sensitive to the variance factor $a$.

bash scripts/benchmark_gaussian.sh

• Sine-type activation functions are sensitive to the frequency factor $\omega$.

# Sine
bash scripts/benchmark_siren.sh

# Incode-Sine
bash scripts/benchmark_incode-sine.sh

Periodic activation functions are sensitive to initialization schemes

bash scripts/benchmark_sensitive_init.sh

Fourier-ASR is parameter-insensitive

When model capacity is limited, larger $\Omega$ in the input layer improve the performance.

bash scripts/benchmark_Fourier_omega.sh

Citation

@article{Li_2025,
   title={Representing Sounds as Neural Amplitude Fields: A Benchmark of Coordinate-MLPs and a Fourier Kolmogorov-Arnold Framework},
   volume={39},
   number={23},
   journal={Proceedings of the AAAI Conference on Artificial Intelligence},
   publisher={AAAI},
   author={Li, Linfei and Zhang, Lin and Wang, Zhong and Zhang, Fengyi and Li, Zelin and Shen, Ying},
   year={2025},
   month=apr, pages={24458–24466} }