Speech intelligibility is a critical aspect of building science, particularly in educational buildings where poor sound quality may have a detrimental impact on students' learning and teachers’ health. However, considering the numerous building regulations proposing varying definitions and ranges of acoustic comfort, calculating the necessary acoustic indicators can be challenging for designers. Speech intelligibility is a crucial component of indoor acoustics and acoustic comfort and can be calculated using formulas, simulation software, and data-based web tools. While formulas are fast, they lack details; acoustic simulation software is highly accurate but time-consuming and expensive. Data-based web tools, including machine learning algorithms, offer both speed and accuracy and are widely accessible. In this study, we present a system utilizing machine learning techniques to predict acoustic indicators, numeric and heatmap, in an educational building. The Pachyderm plugin in the Grasshopper was utilized to conduct extensive simulations in a single educational space, focusing on acoustic indicators in six different frequencies and general modes. Using Catboost and the pix2pix algorithm, the prediction models provide numerical and image indices on the developed dataset. Also, SHAP values were employed to interpret the Catboost model, analyzing the significance of each feature. The results showed remarkable accuracy, (i.e., between 89 % and 99 %) in the numerical portion, and PSNR index ranging from 0.817 to 0.970, and an SSIM index ranging from 15.56 to 31.57 in the image section. By utilizing data-driven methods, the system provides an efficient and accurate approach to calculating acoustic indicators, helping to ensure optimal acoustic environment in educational buildings.

Achieving optimal speech intelligibility in educational settings is crucial for effective learning.
Designers face challenges due to the diversity of building regulations, which define acoustic
comfort in various ways. Objective acoustic parameters such as Definition (D50) and the
Speech Transmission Index (STI) are pivotal in assessing acoustic quality tailored to a room's
function, with STI being especially indicative of speech intelligibility. To address the need for
quick, accurate predictions of D50 and STI values across classroom areas, this research
employs a surrogate machine learning (ML) approach. Our methodology involves simulating
acoustic properties in a single educational room using the Pachyderm plugin within
Grasshopper to analyze D50 at three key frequencies (125, 1000, and 4000 Hz) and the overall
STI. We utilize the CatBoost algorithm as a surrogate model to predict the acoustic performance
of individual sensors. The effectiveness of our model is assessed using the R2 score, Mean
Absolute Error (MAE), and Mean Squared Error (MSE) for individual sensors, along with
Pearson correlation for comprehensive sensor analysis. The results demonstrate the high
performance and potential of this surrogate ML approach in generating detailed and accurate
acoustic heatmaps, thus ensuring enhanced acoustic comfort in educational environments. This
method provides a cost-effective and efficient solution for real-time acoustic assessment, paving
the way for improved educational building design.