Reconstruction of three‐dimensional sound field system based on machine learning method

Abstract The reconstruction of three‐dimensional (3D) audio systems is essential for enhancing immersive experiences and advancing film and television technology. However, the 22.2 multichannel 3D playback system requires precise speaker arrangement and a specific playback environment, and its computational demands hinder its widespread adoption. This paper addresses these challenges by calculating the ideal 22.2 system sound signal based on the actual playback environment and developing a neural network model for speaker signal allocation using a sound signal dataset. Trained with TensorFlow, the model determines the allocation weights for each speaker, optimizing signal distribution. Experimental comparisons show that our model reduces the mean square error and computation time by 25.27%, significantly improving computational efficiency and error performance over the original system, thereby facilitating the efficient reconstruction of 3D audio for diverse playback environments.