yoyololicon / hrtf-ilp

HRTF time-of-arrival and phase estimation.

Time-of-arrival estimation and phase unwrapping of HRTF with integer linear programming

Source code of the papers Time-of-arrival estimation and phase unwrapping of head-related transfer functions with integer linear programming (AES Convention 156) and Arbitrarily sampled signal reconstruction using relative difference features (GSP Workshop 2024).

Getting started

Please install the required packages by running the following command. You'll need to have Python >=3.11 installed on your system.

pip install -r requirements.txt

Preprocessing

Given a single HRTF sofa file, preprocess.py will compute its TOAs using 36 different configurations we stated in the paper and store the results into a folder.

python preprocess.py input.sofa output_folder --toa-weight 0.1 --oversampling 10

Here, --toa-weight controls the value of $\lambda$ and $w_{\delta, i}$ of exponential weighting in the paper. The output npz files has the format of {edgeslist/ilp/l2}_toa_{True/False}_cross_{True/False}_{angle/dot/none}.npz. They corresponds to the terms in the paper as follows:

Name	Name in ther paper
edgeslist	EDGY
ilp	SIMP
l2	LS
toa_True	w/ Min.
toa_False	w/o Min.
cross_True	w/ Cross
cross_False	w/o Cross
angle	EXP
dot	CORR
none	NONE

Experiments

ITD/aligned-HRIRs reconstruction

Please create a folder named processed and put the preprocessed directories in it. Then, execute the notebook eval to evaluate the estimated TOAs.

Noise robustness

Please execute the notebook noise-robust to see the results. The notebook includes the code to generate Figure 2 in the paper. Changing the sofa file to the one you want to test will generate the results for the given HRTF.

Phase unwrapping

Please execute the notebook phase-unwrapping to see the results. The notebook includes the code to generate Figure 4 and 5 in the paper. The notebook uses kamui for ILP-based unwrapping.

Citation

Coming soon.