Deep-Learning-Driven Sensory Profiling of Headphone Target Curves with Adaptive Listening Test Validation

Ravizza G, Villegas J, Stegenborg-Andersen T, Sarria M. GM. Deep-Learning-Driven Sensory Profiling of Headphone Target Curves with Adaptive Listening Test Validation. In: Express Paper. Audio Engineering Society; 2026. Paper 432. Available from: https://aes.org/publications/elibrary-page/?id=23157

@inproceedings{Ravizza2026_23157,
  author = {Ravizza, Gabriele and Villegas, Julián and Stegenborg-Andersen, Tore and Sarria M., Gerardo M.},
  title = {{Deep-Learning-Driven Sensory Profiling of Headphone Target Curves with Adaptive Listening Test Validation}},
  note = {Paper 432},
  year = {2026},
  month = may,
  publisher = {Audio Engineering Society},
  url = {https://aes.org/publications/elibrary-page/?id=23157}
}

TY  - CPAPER
TI  - Deep-Learning-Driven Sensory Profiling of Headphone Target Curves with Adaptive Listening Test Validation
AU  - Ravizza, Gabriele
AU  - Villegas, Julián
AU  - Stegenborg-Andersen, Tore
AU  - Sarria M., Gerardo M.
M1  - Paper 432
PY  - 2026
DA  - 2026/05/28
UR  - https://aes.org/publications/elibrary-page/?id=23157
PB  - Audio Engineering Society
LA  - en
AB  - Identifying headphone target curves based on perceptual data from a target group of untrained listeners is challenging mainly because of the primary characteristic of the untrained listener (the untrainedness). This characteristic limits the perceptual tasks they can reliably perform and the quality of the perceptual data collected regardless of any care taken in the test design to ensure data quality. When working with preference, the issue is compounded as there is no guarantee [1] that all untrained listeners have the same desired frequency target curve when listening to headphones. In previous work by the authors [2], a method to reliably measure a preference target curve for headphones on a B&K HATS 5128 using untrained consumers was investigated. In this work, the findings from the previous paper were applied to a larger cohort of listeners, yielding an updated target curve. In the listening tests performed during this study, candidate target curves were generated using an Interactive Differential Evolution (IDE) listening experiment [2] from three culturally distinct listener panels (Denmark, Japan, and Colombia), enabling continuous exploration of the perceptually relevant tuning space while reducing cognitive load. It was found that although the overall level of gains was similar across sites, the preferred gains at some bands differed among groups. Compared with a previously identified curve, the three groups preferred lower gains at high frequencies, especially at 4 and 8 kHz. In addition, the converged gain sets (headphone target curves) were analyzed using a Virtual Listener Panel (VLP), which is a Deep Learning model trained on large-scale expert evaluations to predict perceptual attributes from rendered musical material. Predicted attributes indicated minor differences among the tested curves, with bass and treble strength attributes yielding the most divergent predicted contributions to the target profiles.
ER  -