Personalized Audio Systems—A Bayesian Approach

Nielsen JB, Jensen BS, Hansen TJ, Larsen J. Personalized Audio Systems—A Bayesian Approach. In: AES Convention 135. Audio Engineering Society; 2013.pp. 9000:1-. Available from: https://aes.org/publications/elibrary-page/?id=17048

@inproceedings{Nielsen2013_17048,
  author = {Nielsen, Jens Brehm and Jensen, Bjørn Sand and Hansen, Toke Jansen and Larsen, Jan},
  title = {{Personalized Audio Systems—A Bayesian Approach}},
  booktitle = {AES Convention 135},
  pages = {9000:1-},
  year = {2013},
  month = oct,
  publisher = {Audio Engineering Society},
  url = {https://aes.org/publications/elibrary-page/?id=17048}
}

TY  - CPAPER
TI  - Personalized Audio Systems—A Bayesian Approach
AU  - Nielsen, Jens Brehm
AU  - Jensen, Bjørn Sand
AU  - Hansen, Toke Jansen
AU  - Larsen, Jan
T2  - AES Convention 135
SP  - 1
M1  - Paper 9000
PY  - 2013
DA  - 2013/10/06
UR  - https://aes.org/publications/elibrary-page/?id=17048
PB  - Audio Engineering Society
LA  - en
AB  - Modern audio systems are typically equipped with several user-adjustable parameters unfamiliar to most listeners. To obtain the best possible system setting, the listener is forced into non-trivial multi-parameter optimization with respect to the listener's own objective and preference. To address this, the present paper presents a general interactive framework for robust personalization of such audio systems. The framework builds on Bayesian Gaussian process regression in which the belief about the user's objective function is updated sequentially. The parameter setting to be evaluated in a given trial is carefully selected by sequential experimental design based on the belief. A Gaussian process model is proposed that incorporates assumed correlation among particular parameters, which provides better modeling capabilities compared to a standard model. A five-band constant-Q equalizer is considered for demonstration purposes, in which the equalizer parameters are optimized for each individual using the proposed framework. Twelve test subjects obtain a personalized setting with the framework, and these settings are significantly preferred to those obtained with random experimentation.
ER  -