whojavumusic/hifi_harp

HiFi-HARP: High-Fidelity Hybrid Ambisonic Room Impulse Responses (FOA Version)

Dataset Summary

HiFi-HARP (FOA Version) is a large-scale dataset of First-Order Ambisonic (FOA) Room Impulse Responses (RIRs) generated via high-fidelity hybrid acoustic simulation. It is derived from the larger 7th-order HiFi-HARP dataset.

This dataset is designed to support research in spatial audio, specifically for tasks such as:

• Sound Source Localization (SSL)
• Dereverberation
• Acoustic Parameter Estimation ($T_{60}$, DRR)
• Spatial RIR Generation

The RIRs are simulated in realistically furnished indoor scenes from the 3D-FRONT dataset, ensuring geometric and semantic complexity.

Dataset Structure

Due to storage constraints, this repository hosts the First-Order Ambisonics (FOA) version of the dataset. The data is partitioned into two compressed archives:

• FOA_train.zip: Contains the training set of FOA-RIRs.
• FOA_valid.zip: Contains the validation/test set of FOA-RIRs.

File Format

• Audio Format: .wav files
• Channels: 4 channels.
• Sample Rate: 16 kHz.

Simulation Methodology

The RIRs were generated using a robust Hybrid Simulation Pipeline that combines wave-based and geometric acoustic methods to achieve high fidelity across the frequency spectrum:

1. Low Frequencies (< 900 Hz): Simulated using a Finite-Difference Time-Domain (FDTD) wave solver to accurately model diffraction and room modes.
2. High Frequencies (> 900 Hz): Simulated using a geometric ray-tracing approach to capture specular reflections and scattering.
3. Scene Data: Furnished rooms from the 3D-FRONT dataset with varying volumes, shapes, and material properties.

Note: While the original dataset was generated at 7th-order (64 channels), this version has been truncated to 1st-order (4 channels) for ease of distribution.

Citation

@article{saini2025hifi, title={HiFi-HARP: A High-Fidelity 7th-Order Ambisonic Room Impulse Response Dataset}, author={Saini, Shivam and Peissig, J{"u}rgen}, journal={arXiv preprint arXiv:2510.21257}, year={2025} }