Get Raw Room Spatial Impulse Response Ray Tracing Data Accepted at AES 2025 Europe

(under active development)

Paper Link

• 12/19/2025: We added NVIDIA OptiX support. Try it out!

• 07/13/2025: We updated the auralization logic, which now allows for any higher-order ambisonic generation, as well as an improved logic for generating the ambisonic signals. Check it out!

• 07/10/2025: There was a critical bug in both auralizer and the ray generator. The ray generator does not output listener directions in the output SIR, and the auralizer was using the listener position rather than their directions for calculation. Thanks to @FEAfeatherTHER for spotting this bug!

Important licensing information: This repository is developed on the basis of GSound and PyGSound, which follow their respective license. See license for details.

We will gradually remove all depenedencies to PyGSound. All code added additionally are open-sourced under Apache 2.0 license.

Special thanks to Dr. Carl Schissler and Dr. Zhenyu Tang for very helpful email correspondance, and thanks to Xuzhou Ye and Junjie Shi for incredibly fruitful discussions that guided this project's direction.

conda create -n gsoundsir python=3.10
conda activate gsoundsir
cd auralizer
pip install .
cd ../ray_generator
pip install .

GSound-SIR is a Python-based extension of GSound designed for efficient room acoustics simulation and high-order Ambisonic impulse response (SIR) generation. While GSound provides a powerful C++ ray-tracing core for interactive applications, it traditionally acts as a black box—limiting users to final acoustic results. GSound-SIR changes that by exposing raw ray-level data, enabling in-depth inspection, custom analysis, and advanced post-processing.

Key Highlights:

• Raw Ray Data Access: Directly inspect propagation paths (arrival times, directions, energy, etc.) in Python.
• Higher-Order Ambisonics: Synthesize spatial impulse responses up to 9th-order Ambisonics.
• Optimized Data I/O: Efficiently store ray data in Parquet format for seamless integration with modern data analysis toolchains.
• Energy-Based Filtering: Keep only the most acoustically significant ray paths (e.g., top X% by energy) to reduce storage overhead.
• Pythonic Interface: Built using pybind11 for easy integration into Python workflows.

Ray tracing is a widely-used technique in acoustic simulations and spatial audio research. It strikes a balance between:

• Accuracy: Capturing complex reflection/diffraction.
• Efficiency: Scaling to large or arbitrarily shaped environments more feasibly than many finite-difference methods.

Traditional open-source libraries often output final impulse responses (IRs) without providing access to intermediate ray-level data. GSound-SIR bridges this gap, allowing full inspection of each ray path, enabling advanced custom analysis and flexible post-processing.

• Decoupled Ray Generation and Auralization
  Run the core C++ ray tracer to extract raw ray data, then process or auralize in Python.
• Energy-Based Filtering
  Output only the top-X or top-X% of highest-energy rays, reducing storage needs while preserving accuracy.
• Parquet Output
  Export large datasets in columnar format for efficient reading and analysis in Python (pandas, PySpark, etc.).
• Flexible Ambisonic Orders
  Convert raw rays into high-order Ambisonic impulse responses (up to 9th order).
• Extensible
  Implement your own custom weighting, filtering, or HRTF-based rendering, and efficiently call them in Python (See auralizer/src/cpp/binding.cpp)

Empirical results demonstrate:

• Linear Scaling with Ray Count: Computation time grows linearly with diffuse + specular ray count.

• Energy Concentration: A small fraction of rays can carry the majority of acoustic energy (top 0.1% ~ 80+% energy in certain tests).

• Disk I/O: Writing large volumes of ray data is often the bottleneck. Parquet storage plus energy-based filtering can drastically reduce size and time.

• Stationary Sources: Currently optimized for stationary sources. Moving source support is on our roadmap.
• Ambisonic Auralization: Extensible to other rendering methods (e.g., binaural/HRTF-based) with custom Python scripts.
• Potential for Deep Learning: The concentrated nature of ray energy suggests that neural upsampling or ML-based reflection modeling is a promising direction.

If you use GSound-SIR in your work, we appreciate citations:

@misc{zang2025gsoundsirspatialimpulseresponse,
      title={GSound-SIR: A Spatial Impulse Response Ray-Tracing and High-order Ambisonic Auralization Python Toolkit}, 
      author={Yongyi Zang and Qiuqiang Kong},
      year={2025},
      eprint={2503.17866},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2503.17866}, 
}

You may also reference the underlying GSound project where appropriate.

• GSound-SIR Python Bindings and Extensions: Apache License 2.0
• Core GSound: Distributed under its own license.

Please review both licenses before integrating GSound-SIR into your projects.

Thank you for using GSound-SIR! Feel free to open issues for bugs or feature requests, and we welcome pull requests. For deeper discussions or collaboration inquiries, please contact Yongyi Zang ([email protected]).