Towards real-time two-dimensional wave propagation for articulatory speech synthesis

171st Meeting of the Acoustical Society of America |

Published by Acoustical Society of America

View Publication | Download PDF

The precise simulation of voice production is a challenging task, often characterized by a tradeoff between quality and speed. The usage of 3D acoustic models of realistic vocal tracts produces extremely precise results, at the cost of running simulations that may take several minutes to synthesize a few milliseconds of audio. In contrast, 1D articulatory vocal synthesizers rely on highly simplified acoustic and anatomical models to achieve real-time performances, but can only partially match the spectra of realistic vocal tracts. In this work, we present a novel articulatory vocal synthesizer, based on a fast 2D propagation model running on a graphics card (GPU). The system can run in real-time under specific conditions and, differently from 1D synthesizers, allows for simulating airflow propagation through asymmetric and curved geometries. This paper covers details on the GPU implementation of the different components of the system, including the 2D Finite-Difference Time-Domain wave solver and the excitation mechanism. A preliminary evaluation is presented,
using area functions to simulate static vowels. Three different resolutions are tested, combined with
two alternative ways of discretizing the 2D geometries. The computed formants are overall characterized by small positional errors while computational times are comparable with those from 1D systems.