Modeling Time-Varying Nonlinear Loudspeaker Displacement With Physics-Guided Dilated Convolutional Networks

Su F, Bian Z, Zhu R, Shen Y. Modeling Time-Varying Nonlinear Loudspeaker Displacement With Physics-Guided Dilated Convolutional Networks. J Audio Eng Soc. 2026;74(11):839-851. doi:10.17743/jaes.2026.0306

@article{Su2026_23503,
  author = {Su, Fenshan and Bian, Zitong and Zhu, Rui and Shen, Yong},
  title = {{Modeling Time-Varying Nonlinear Loudspeaker Displacement With Physics-Guided Dilated Convolutional Networks}},
  journal = {Journal of the Audio Engineering Society},
  volume = {74},
  number = {11},
  pages = {839--851},
  url = {https://doi.org/10.17743/jaes.2026.0306},
  doi = {10.17743/jaes.2026.0306},
  year = {2026},
  month = nov,
  publisher = {Audio Engineering Society}
}

TY  - JOUR
TI  - Modeling Time-Varying Nonlinear Loudspeaker Displacement With Physics-Guided Dilated Convolutional Networks
AU  - Su, Fenshan
AU  - Bian, Zitong
AU  - Zhu, Rui
AU  - Shen, Yong
T2  - Journal of the Audio Engineering Society
J2  - J. Audio Eng. Soc.
VL  - 74
IS  - 11
SP  - 839
EP  - 851
PY  - 2026
DA  - 2026/11/10
DO  - 10.17743/jaes.2026.0306
UR  - https://doi.org/10.17743/jaes.2026.0306
PB  - Audio Engineering Society
LA  - en
AB  - Moving-coil loudspeakers exhibit strong large-signal nonlinearities and complex time-varying mechanical drift due to suspension viscoelasticity, which challenges accurate long-term displacement prediction. This paper proposes a physics-guided WaveNet (PG-WaveNet) that integrates physical priors as multichannel local conditioning. The conditioning signals are built from low-frequency voice-coil current energy and leaky-integral histories. Receptive field analysis shows that strongly time-varying behavior requires temporal context extending well below resonance, where purely data-driven WaveNet models become limited. Experiments on two loudspeakers with different degrees of viscoelasticity show that PG-WaveNet improves both accuracy and robustness over a standard WaveNet, a long short-term memory network, and a nonlinear state-space function model. On the strongly time-varying dataset, PG-WaveNet achieves 2.6 dB lower NMSE and 4.75 dB lower maximum absolute error than the standard WaveNet and maintains stable tracking over continuous 60-s intervals under multitone and music excitation. Ablation results confirm that medium-term and long-term conditioning channels are essential when mechanical drift exceeds the network’s receptive field.
ER  -