Estimating room acoustic descriptors from bag-of-vectors representation with transformers

In this paper, we propose a novel deep learning method for room acoustic descriptor estimation. Certain descriptors are highly important in assessing acoustic quality, therefore estimating them during the planning phase is a crucial part of designing indoor spaces. Traditional approaches rely on either computationally expensive numerical methods, or statistical formulae with insufficient accuracy. Our solution is FRAPPE (fast room acoustic prediction and parameter estimation), which applies lightweight transformer-based neural networks to estimate acoustic descriptors in rectangular rooms, utilizing a “bag-of-vectors” representation that is capable of capturing diverse interior designs. We employ transformers without positional encoding, highlighting the broad applicability of the architecture outside of traditional domains. FRAPPE achieves high accuracy and operates at near-instant speed, providing a better cost–accuracy balance than either ray tracing methods or empirical formulae. It is the first transformer-based approach that is applicable in the design phase, and it offers a more general deep learning solution for acoustic descriptor estimation than any prior methods. The accuracy, inference speed and versatility of FRAPPE makes it a valuable innovation for architectural design, supporting better decisions during the early stages of room planning.