During the initial phases of design, engineers and architects lack quick and accurate ways to evaluate acoustic performance, instead depending on laborious full-scale simulations or simplified, often imprecise formulas. The goal of this study is to create a machine-learning framework that can forecast important acoustic performance metrics from various spatial configurations and sound absorption patterns. A dataset of room geometries (dimensions, shape parameters, geometric descriptors) and absorption coefficients (walls, ceiling, floor) has been generated via a digital design pipeline, and acoustic simulations have been executed to obtain target metrics (e.g., sound pressure level, reverberation time, and clarity indices). This dataset was then used to train a neural network that could be integrated into early-stage design workflows, enabling geometry tweaks and material decisions to be evaluated acoustically in real time. This presentation reports mid-term findings, including dataset composition and training strategy, and discusses initial prediction accuracy.