Pitch-excited ARMA lattice model for speech synthesis

Abstract

We present the ARMA lattice model for speech synthesis. By adopting a pole-zero approach, this model overcomes the limitation of the absence of zeros in the LPC model, which is based entirely on an all-pole AR model for the representation of the vocal tract. Therefore, a more natural synthesized speech can be achieved using this ARMA approach, especially in speech signals that contain nasal, fricative and plosive sounds. The structure of the ARMA lattice model is numerically stable, which is desirable for robust speech synthesis applications. Simulation results indicate that the quality of the synthesized speech of spoken sentences is encouraging, and a better result can be achieved by optimizing the parameters to be used.