Speech Synthesis Method for Dialogue 
and Psychological Assessment of Synthetic Speech
<P>
Hideki Kasuya, Wen Ding, Jin-lin Lu, Naoki Koshikawa 
and Yoshihisa Watanabe
<P>
Department of Electrical and Electronic Engineering, Utsunomiya University<P>
2753 Ishii-machi, Utsunomiya 321, Japan<P>
e-mail: kasuya@utsunomiya-u.ac.jp
<P>
  In order to develop a speech synthesis method for a dialogue system
and to derive appropriate methods to make assessment of synthetic 
speech in the dialogue environment, three major efforts have been made 
this year: Exploring a novel method to extract the parameters  
of a voice source and ARMA model, (2)developing a speaking rate 
conversion system, and (3)building up an integrated speech analysis
software system.

  A simulated annealing method was successfully used to estimate
the voice source parameters of a Rosenberg-Klatt model, as it 
was applied together with a nonstationary Kalman filter model 
to estimate ARMA parameters.  This method was applied to the analysis
of some of the VCV(vowel-consonant-vowel) sound segments 
uttered in a sentence context of [sorewa VCV desuka?] (Is it VCV ?).  
The ARMA parameters together with the voice source parameters 
analyzed from each of the VCV segments were concatenated 
to produce a meaningful sentence, /imagamigoronoume/
(Plum flowers are at their best.).  The pitch contour
was created from the same sentence uttered separately.  The synthetic
sound proved to be natural sounding and quite intelligible.  Our new method
seems promising in developing a speech synthesizer for dialogue system.

  Rhythm and tempo of speech seem to play an important role in the
inter-speaker communication.  To investigate psychological
significance of the rhythm and tempo in the dialogue, we have to have
an experimental apparatus to control in various ways rhythm and tempo
of a dialogue speech.  We have developed a semi-automatic speaking
rate conversion system based on a repetition and elimination of a
single period waveform.

  An integrated speech analysis software has been made to perform 
efficiently the measurement and statistical analysis of prosodic 
features in dialogue speech.  This system is to be used next year
to gain acoustic phonetic knowledge of the dialogue speech.<P>