First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

A Parametric Model of Speech Signals: Application to High Quality Speech Synthesis by Spectral and Prosodic Modifications

Thierry Galas, Xavier Rodet

LAFORIA, UA CNRS N1095, Paris, France

We propose here a new parametric model and its application to speech synthesis. In our source-filter model, the source is described by spectro temporal events. The filter combine an all-pole filter model for the vocal tract and a de-emphasis filter corresponding to the lip radiation and glottal spectrum slope. Source events are singular or belong to a continuum or pseudo-continuum of events. Examples of singular events are the burst of noise at release of plosive or isolated glottal pulses. Pseudo-continua of events are quasi periodic glottal pulses with their intrinsic irregularities with possible superimposed fricative noise, or pure noise signals as in unvoiced fricatives. Our model allows for a precise and perceptually satisfying description of speech signal and simultaneously provides more flexibility for prosodic modifications. We present an analysis method according to our model using any spectral estimation technique such as AR or homomorphic estimations. We also present an overlap-add synthesis method using the analysis data. We show that our method can be interpreted in terms of frequency domain spectral interpolation as an ARMA model.

Full Paper

Bibliographic reference.  Galas, Thierry / Rodet, Xavier (1990): "A parametric model of speech signals: application to high quality speech synthesis by spectral and prosodic modifications", In ICSLP-1990, 801-804.