Researchers at the MIT-IBM Watson AI Lab and MIT are experimenting with spatial acoustic information. They plan to use sound to model physical spaces with a new machine-learning model. The ML model can simulate how sound will propagate within a room and how listeners perceive it at different locations.
The researchers used a model similar to the Implicit Neural representation model (used to generate smooth 3D reconstructions) and captured how sound waves travel through space. The system will model spatial acoustics and then learn the underlying geometry of the space. With this information, the model builds visual renderings of the space, similar to how humans use sounds to estimate their physical environment.
Until now, researchers have only modeled vision using the property of photometric consistency. But with sound, this property is inconsistent as changing locations also changes how sound is perceived.
To overcome such limitations in acoustic modeling, the researchers focused on two properties: sound’s reciprocal nature and the impact of geometric objects. To incorporate these, the model uses Neural Acoustic Fields (NAFs). NAFs are grid-based neural networks that capture architectural features of space.
Researchers can provide the NAF with visual data about a scene and a few spectrograms that illustrate how audio might sound when the emitter and listener are situated at specific points around the room.
The new technique of mapping physical spaces using sound will open opportunities for “immersive multimodal experiences in metaverse applications.” For more detailed information, refer to the research paper “Learning Neural Acoustic Fields.”