NAT: Neural Acoustic Transfer for Interactive Scenes in Real Time

Area of Science:

• Acoustics
• Computer Science
• Signal Processing

Background:

• Traditional acoustic transfer methods require extensive precomputation and struggle with dynamic scene changes.
• Representing and rendering fluctuating acoustic transfer distributions in real-time is challenging.

Purpose of the Study:

• To introduce Neural Acoustic Transfer, a novel method for real-time prediction of dynamically evolving sound fields.
• To enable efficient and accurate modeling of sound behavior in complex, changing acoustic environments.

Main Methods:

• Leveraging implicit neural representations to encode acoustic transfer functions and their variations.
• Developing a fast Monte-Carlo-based boundary element method (BEM) approximation for generating high-quality training data.
• Implementing strategies to mitigate singularities during training data synthesis.

Main Results:

• Achieved real-time prediction of sound fields with runtime efficiency within milliseconds for 30-second audio.
• Demonstrated numerical accuracy and efficiency through validation in diverse acoustic transfer scenarios.
• Successfully modeled complex sound radiation spaces and dynamic environmental interactions.

Conclusions:

• Neural Acoustic Transfer effectively models sound behavior in dynamic environments, overcoming limitations of prior methods.
• The approach offers efficient and accurate solutions for interactive applications like virtual and augmented reality.
• Implicit neural representations combined with BEM approximation provide a robust framework for real-time acoustic modeling.