Deep Neural Network Based HRTF Personalization Using Anthropometric Measurements

Chun CJ, Moon JM, Lee GW, Kim NK, Kim HK. Deep Neural Network Based HRTF Personalization Using Anthropometric Measurements. In: AES Convention 143. Audio Engineering Society; 2017. Paper 9860. Available from: https://aes.org/publications/elibrary-page/?id=19257

@inproceedings{Chun2017_19257,
  author = {Chun, Chan Jun and Moon, Jung Min and Lee, Geon Woo and Kim, Nam Kyun and Kim, Hong Kook},
  title = {{Deep Neural Network Based HRTF Personalization Using Anthropometric Measurements}},
  booktitle = {AES Convention 143},
  note = {Paper 9860},
  year = {2017},
  month = oct,
  publisher = {Audio Engineering Society},
  url = {https://aes.org/publications/elibrary-page/?id=19257}
}

TY  - CPAPER
TI  - Deep Neural Network Based HRTF Personalization Using Anthropometric Measurements
AU  - Chun, Chan Jun
AU  - Moon, Jung Min
AU  - Lee, Geon Woo
AU  - Kim, Nam Kyun
AU  - Kim, Hong Kook
T2  - AES Convention 143
M1  - Paper 9860
PY  - 2017
DA  - 2017/10/06
UR  - https://aes.org/publications/elibrary-page/?id=19257
PB  - Audio Engineering Society
LA  - en
AB  - A head-related transfer function (HRTF) is a very simple and powerful tool for producing spatial sound by filtering monaural sound. It represents the effects of the head, body, and pinna as well as the pathway from a given source position to a listener’s ears. Unfortunately, while the characteristics of HRTF differ slightly from person to person, it is usual to use the HRIR that is averaged over all the subjects. In addition, it is difficult to measure individual HRTFs for all horizontal and vertical directions. Thus, this paper proposes a deep neural network (DNN)-based HRTF personalization method using anthropometric measurements. To this end, the CIPIC HRTF database, which is a public domain database of HRTF measurements, is analyzed to generate a DNN model for HRTF personalization. The input features for the DNN are taken as the anthropometric measurements, including the head, torso, and pinna information. Additionally, the output labels are taken as the head-related impulse response (HRIR) samples of a left ear. The performance of the proposed method is evaluated by computing the root-mean-square error (RMSE) and log-spectral distortion (LSD) between the referenced HRIR and the estimated one by the proposed method. Consequently, it is shown that the RMSE and LSD for the estimated HRIR are smaller than those of the HRIR averaged over all the subjects from the CIPIC HRTF database.
ER  -