Exploring Frequency-Domain Feature Modeling for HRTF Magnitude Upsampling

Official PyTorch implementation of the paper:

Exploring Frequency-Domain Feature Modeling for HRTF Magnitude Upsampling

This repository investigates different frequency-domain modeling strategies (MLP, Conv1D, Dilated Conv, Conformer, etc.) for sparse-to-dense HRTF magnitude upsampling on the SONICOM dataset.

---

1. Dataset Preparation

We use the SONICOM HRTF dataset (200 subjects, 793 directions).

---

Option A — Download raw SOFA files and preprocess

Step 1: Download SONICOM dataset

You can batch-download .sofa files using:

python tools/download_sonicom.py

This will download files:

raw/P0001.sofa
...
raw/P0200.sofa

---

Step 2: Preprocess to generate sonicom.npz

python tools/preprocess_sonicom.py

This will generate:

data/sonicom.npz

The saved file contains:

H_log: (200, 793, 2, 106)

• 200 subjects
• 793 directions
• 2 ears
• 106 frequency bins
• Log-magnitude in dB

---

Option B — Directly use provided .npz

data/sonicom.npz

You can skip the raw download and preprocessing steps.

---

2. Training

Training is controlled by a YAML configuration file.

Edit:

configs/default.yaml

Set the number of sparse measurements:

data:
  n_measurements: 3   # choose from {3, 5, 19, 100}

---

Run training

python src/train.py --config configs/default.yaml

Training will:

• Load SONICOM dataset
• Split subjects into train/test
• Train model
• Save best model to:

runs/sonicom_<variant>_<date>/best_model.pt

---

3. Evaluation

After training, evaluate using:

python src/eval.py \
    --config configs/default.yaml \
    --ckpt runs/<run_name>/best_model.pt

Evaluation reports:

• LSD (Log Spectral Distortion) on unseen directions
• Broadband ILD MAE

Example output:

Evaluation Results
==============================
Sparse measurements : 3
Train LSD (unseen)  : 4.12 dB
Test  LSD (unseen)  : 4.85 dB
Train ILD MAE       : 1.24 dB
Test  ILD MAE       : 1.37 dB
==============================

---

4. Model Variants

Set model variant in YAML:

model:
  variant: conformer   # options:
                       # mlp
                       # conv
                       # dilated_b2
                       # conformer
                       # conformer_wo_conv

All variants share the same architecture except for the frequency modeling block.

---

5. Environment

Tested with:

Python 3.10
PyTorch 2.6.0

Install dependencies:

pip install -r requirements.txt

Install PyTorch (recommended)

For CUDA 12.4:

conda install pytorch=2.6.0 pytorch-cuda=12.4 -c pytorch -c nvidia

---

6. Citation

If you use this code, please cite:

@article{chen2026exploring,
  title={Exploring Frequency-Domain Feature Modeling for HRTF Magnitude Upsampling},
  author={Chen, Xingyu and Bi, Hanwen and Ma, Fei and Zhao, Sipei and Cheng, Eva and Burnett, Ian S},
  journal={arXiv preprint arXiv:2602.11670},
  year={2026}
}