Real-Time Adaptive Parametric Equalization for Joint Room Response Compensation and Tonal Shape Reproduction

Ahn Y. Real-Time Adaptive Parametric Equalization for Joint Room Response Compensation and Tonal Shape Reproduction. J Audio Eng Soc. 2026;74(11):904-918. doi:10.17743/jaes.2026.0314

@article{Ahn2026_23508,
  author = {Ahn, Yongho},
  title = {{Real-Time Adaptive Parametric Equalization for Joint Room Response Compensation and Tonal Shape Reproduction}},
  journal = {Journal of the Audio Engineering Society},
  volume = {74},
  number = {11},
  pages = {904--918},
  url = {https://doi.org/10.17743/jaes.2026.0314},
  doi = {10.17743/jaes.2026.0314},
  year = {2026},
  month = nov,
  publisher = {Audio Engineering Society}
}

TY  - JOUR
TI  - Real-Time Adaptive Parametric Equalization for Joint Room Response Compensation and Tonal Shape Reproduction
AU  - Ahn, Yongho
T2  - Journal of the Audio Engineering Society
J2  - J. Audio Eng. Soc.
VL  - 74
IS  - 11
SP  - 904
EP  - 918
PY  - 2026
DA  - 2026/11/10
DO  - 10.17743/jaes.2026.0314
UR  - https://doi.org/10.17743/jaes.2026.0314
PB  - Audio Engineering Society
LA  - en
AB  - In an ordinary listening room, the room itself colors the reproduced signal, although the user may want a nonflat tonal balance. Room response and equalization add on a decibel scale, so an EQ must undo the room and realize the target jointly, not in two independent stages. This engineering report describes a real-time adaptive parametric EQ that performs both in a single feed-forward network; the task is objective reproduction of a prescribed spectral target under room coloration, not a subjective-preference problem. A compact convolutional encoder with separate room-response and preference branches feeds a two-stage head that predicts a room-correction prior, then outputs a small set of biquad parametric EQ parameters whose magnitude response is rendered exactly, in closed form, during training, so the output drives an embedded biquad engine directly without postinference fitting. Trained on synthetic room impulse responses and evaluated on held-out synthetic and real-world (BUT ReverbDB, OpenAIR) sets, it is compared against nonlearned, learned, and stronger dense-output baselines. The dense variants are more accurate in distribution, but under the same parametric output constraint, their advantage falls below a descriptive EQ-peak just-noticeable-difference reference on most samples, whereas the proposed model fits the tested microcontroller in 32-bit floating point (FP32) without quantization.
ER  -