@techreport{oai:ipsj.ixsq.nii.ac.jp:00216634,
 author = {伊藤, 悠貴 and 中村, 友彦 and 小山, 翔一 and 猿渡, 洋 and Yuki, Ito and Tomohiko, Nakamura and Shoichi, Koyama and Hiroshi, Saruwatari},
 issue = {33},
 month = {Feb},
 note = {臨場感のあるバイノーラル信号を合成するためには受聴者本人の頭部伝達関数（head-related transfer function: HRTF）を用いることが望ましい．HRTF の計測には時間がかかるため，少数の観測から HRTF を補間できればより簡便な測定で済み利便性が向上する．従来の球波動関数展開による HRTF 補間方法では任意音源位置の HRTF を簡便に補間できるものの，観測点が少なくなるに従い補間性能が低下する傾向にあった．そこで本稿では，球波動関数展開による HRTF の表現方法とメタ学習を組み合わせた，深層学習に基づく少数観測点からの HRTF 補間手法を提案する．メタ学習では少数観測点から補間を行う状況を模倣して深層ニューラルネットワークを訓練するため，提案法は観測点数が少ない場合でも安定して動作できる．HRTF 補間実験により，提案法は観測点数が少ない場合に従来法よりも高精度に補間が可能であることを示した．, In binaural synthesis, listeners’ individual head-related transfer functions (HRTFs) are necessary for highly-immersive spatial audio. Since HRTF measurement is generally time-consuming, it will be helpful if high-resolution HRTFs are interpolated from a small number of HRTFs obtained by a simple measurement procedure. One of the established HRTF interpolation methods is the method based on spherical wavefunction expansion, which allows estimating HRTFs at arbitrary direction and distance in a simple manner; however, its interpolation accuracy deteriorates as the number of measurements decreases. We propose a deep-neural-network (DNN)-based HRTF interpolation method combining the representation using spherical wavefunction expansion and meta-learning . Since meta-learning simulates the process of interpolation from a small number of measurements to learn DNN using training data, the proposed method will stably estimate HRTFs even when the number of measurements is insufficient. Experimental results indicated that the proposed method achieves high interpolation accuracy compared with the current method when the number of measurements is small.},
 title = {球波動関数展開を用いた深層学習による少数測定データからの頭部伝達関数補間},
 year = {2022}
}