Personalized Voice Reconstruction via 3D Craniofacial Recognition, Acoustic Analysis, and Air–Bone Transfer Modeling

This project introduces a pioneering interdisciplinary framework for reconstructing the self-perceived human voice — the way people hear themselves internally — using a combination of 3D craniofacial recognition, biophysical acoustic modeling, and digital signal processing (DSP). Humans perceive their own voice through a hybrid of air conduction (external sound waves transmitted through the air) and bone conduction (vibrations traveling through the skull and tissues). This unique mix shapes an individual’s internal auditory experience, which differs significantly from external recordings. The proposed system mathematically captures this phenomenon through an acoustic transfer function D(f)D(f)D(f), defined as the ratio between bone and air conduction paths. By scanning the user’s head in 3D using LiDAR or depth sensors, the system extracts anatomical features such as skull geometry, sinus cavity volume, and soft-tissue thickness, all of which affect bone conduction characteristics. These parameters are analyzed by an AI model trained to predict the individualized spectral response D(f)D(f)D(f) directly from facial structure data. The resulting transfer function is then implemented as a minimum-phase FIR filter, which transforms an air-conducted voice recording into its self-perceived equivalent in real time. The research integrates finite element acoustic simulation, cepstral minimum-phase reconstruction, and psychoacoustic testing to validate the perceptual accuracy of the reconstructed voice. Comparative experiments demonstrate that the digitally reconstructed self-voice closely matches the subjective auditory experience reported by each individual.