Comparison and Evaluation of Machine Learning Models for Acoustic Optimization in Live Events

¹ Computer Science Department, SINAI, CEATIC, Universidad de Jaén, 23071, Jaén, Spain
² Universidad de Guayaquil, 090514, Guayas, Ecuador
³ Escuela Superior Politécnica del Litoral, 090112, Guayas, Ecuador
⁴ Universidad San Francisco de Quito, 170901, Quito, Ecuador

^* e-mail: mctb0005@red.ujaen.es;mariuxi.toapantab@ug.edu.ec
^** e-mail: hector.toapantab@ug.edu.ec
^*** e-mail: mathtoap@espol.edu.ec
^**** e-mail: ftoapantav@estud.usfq.edu.ec

Abstract

Acoustic quality in live events directly affects audience engagement, yet static calibration methods fail to adapt to real-time environmental and occupancy changes. This study compiles impulse-response data from diverse venues and systematically evaluates linear regression, decision trees, CNN, LSTM, and a CNN–LSTM hybrid for predicting RT60, EDT, D50, C50, and C80. The hybrid model achieved the highest accuracy (RT60 RMSE = 0.097 s, R² = 0.9997) with sub-100 ms inference, enabling real-time deployment. Integrated with IoT sensors tracking temperature, humidity, and audience density, it dynamically adjusts acoustic parameters and outperforms traditional approaches. These results demonstrate the feasibility of ML- and IoT-driven autonomous acoustic control, enhancing sound fidelity, speech intelligibility, and operational efficiency in live-event environments.