Blind Room Acoustic Parameters Estimation Using Mobile Audio Transformer
Shivam Saini, Jürgen Peissig · 2023
For an accurate representation of a virtual sound source in an Augmented Reality environment, it is crucial to understand the acoustic properties of the current room where the source is to be rendered. Reverberation Time (RT60) and Clarity (C50) are two of the most significant parameters that could negatively impact the plausibility of the source when estimated incorrectly. We propose using an audio transformer to estimate these parameters blindly and solely from a single-channel noisy speech signal. Furthermore, to ensure efficiency and minimum computational complexity, the use of an efficient lightweight architecture is proposed for its suitability for mobile-friendly applications. We evaluate the proposed model with regard to its complexity and accuracy against the existing state-of-the-art approaches in unseen and real acoustic settings. Results demonstrate that the proposed model surpasses the traditional CNN models in terms of complexity, size, speed, and precision.