Application of Low-Complexity Neural Equalizer for Adaptive Sound Equalization in Wireless Earbuds.

Bezzola R, Li Y. Application of Low-Complexity Neural Equalizer for Adaptive Sound Equalization in Wireless Earbuds. In: Convention Paper. Audio Engineering Society; 2025. Paper 10247. Available from: https://aes.org/publications/elibrary-page/?id=23091

@inproceedings{Bezzola2025_23091,
  author = {Bezzola, ri and Li, Yuan},
  title = {{Application of Low-Complexity Neural Equalizer for Adaptive Sound Equalization in Wireless Earbuds.}},
  note = {Paper 10247},
  year = {2025},
  month = oct,
  publisher = {Audio Engineering Society},
  url = {https://aes.org/publications/elibrary-page/?id=23091}
}

TY  - CPAPER
TI  - Application of Low-Complexity Neural Equalizer for Adaptive Sound Equalization in Wireless Earbuds.
AU  - Bezzola, ri
AU  - Li, Yuan
M1  - Paper 10247
PY  - 2025
DA  - 2025/10/14
UR  - https://aes.org/publications/elibrary-page/?id=23091
PB  - Audio Engineering Society
LA  - en
AB  - This paper presents an application of a neural equalization method of low complexity for the adaptive equalization of a wireless insert earbud. The task at hand is the calculation of optimal gains for a cascade of six biquadratic peaking filters with fixed center-frequencies and quality factors. Realistic equalization targets are obtained using measurements of the sound pressure response at the eardrum reference point with an earbud inserted into several synthetic ears and varying degrees of fit between the earbud and ear canal. We compare the accuracy of a neural equalizer that is based on a feed-forward neural network to a baseline method that relies on the self-similarity of biquadratic filters in the log-magnitude scale and calculates the optimal gain values using a least-squares approach, and to a neural equalizer trained on random polynomials to estimate coefficients of fully parametric biquad cascades. We show that for the limited number of six biquadratic filters, ten neurons in the hidden layer of the neural equalizer can provide a measurable improvement in equalization performance over the baseline method. In its published form, the fully parametric equalizer has over 5 Million parameters, which is orders of magnitude larger than the tested feedforward neural equalizer with 136 parameters. Yet it is only in the higher octaves of the equalization range that the larger model outperforms the lean model in the earbud equalization task.
ER  -