Global HRTF Interpolation

• This repository provides overall framework for training and evaluating head-related transfer function (HRTF) interpolation systems proposed in 'Global HRTF Interpolation via Learned Affine Transformation of Hyper-conditioned Features'
• Demo sound samples are available in 'here'

Prepare dataset

This process will re-arrange the original Sofa files, and save them in .pkl format. Set default arguments for load_dir and save_dir of prep_hrir in dataset/preprocess.py by "path to your HUTUBS dataset" (which contains HRIRs and Antrhopometric_measures directory) and "path to the save directory", respectively.

python3 dataset/preprocess.py

Interpolation

Reconstruction test using HUTUBS. This will reproduce the results of Table 2. Note that Table 2 of the paper shows the RMSE averaged for five test folds.

python3 interp-hutubs.py --gpu 0 --test_fold 3                   # Table 2, 'All'
python3 interp-hutubs.py --gpu 0 --test_fold 3 --x_constraint 0  # Table 2, 'Fro'
python3 interp-hutubs.py --gpu 0 --test_fold 3 --y_constraint 0  # Table 2, 'Med'
python3 interp-hutubs.py --gpu 0 --test_fold 3 --z_constraint 0  # Table 2, 'Hor'

Interpolation test using FABIAN. This will reproduce the results of Table 3. Note that Table 3 of the paper shows the RMSE averaged for five test folds.

python3 interp-fabian.py --gpu 0 --test_fold 5                   # Table 3, Ours, 'All'
python3 interp-fabian.py --gpu 0 --test_fold 5 --y_constraint 0  # Table 3, Ours, 'Med'
python3 interp-fabian.py --gpu 0 --test_fold 5 --scale_factor 6  # Table 3, Ours (x1/6), 'All'

Super-resolution

To reproduce the results of Figure 4, please see our super-resolution tutorial Colab notebook.

Train

Training procedures can be found under results directory. Specify the path to your HUTUBS data directory for argument --data_dir (should be the same value as save_dir when preprocessing with dataset/preprocess.py).

python3 main.py --gpus 0,1 --train --cnn_layers 5 --condition hyper --in_ch 16 --p_range 0.2 --test_fold 5 --data_dir $path_to_data_dir

Citation

If you find our work helpful, please cite it as below.

@inproceedings{lee2023global,
  title={Global hrtf interpolation via learned affine transformation of hyper-conditioned features},
  author={Lee, Jin Woo and Lee, Sungho and Lee, Kyogu},
  booktitle={ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={1--5},
  year={2023},
  organization={IEEE}
}