1998 年 34 巻 6 号 p. 607-613
Speech synthesis has been studied for about half century since H. Dudley developed “vocoder” in 1939. There are many synthesis methods, and AR model, PARCOR and LSP play an important role today. Among them, formant-bandwidth representation of AR model is said to be most advantageous to information rates. In this paper, we divided a pulse transfer function of AR model to a partial fraction form using Durand-Kerner method and residue formula. The obtained form is easily implemented by parallel connection of complex first order systems (CFOS). Additionally, we have programed the parameter identification algorithm for digital signal processor (DSP), and we have shown the computational efficiency for real-time processing.