Synthesis of voiced sounds using physics-informed neural networks

Recently, physics-informed neural networks (PINNs) have garnered attention for use as a numerical simulation method for inverse analysis, such as property identification. However, studies on PINNs for conducting acoustic analysis are scarce. Thus, this study developed PINNs that performed acoustic analysis of the vocal tract and synthesized voiced sounds. In addition, PINNs were used to identify glottal source waveforms. Consequently, PINNs were demonstrated to be a promising solution for the inverse problem related to speech production.