Third European Conference on Speech Communication and Technology

Berlin, Germany
September 22-25, 1993


An Analysis of the Performances of the MBE Model When Used in the Context of a Text-To-Speech System

Thierry Dutoit, Henri Leich

Faculte Polytechnique de Mons, Mons, Belgique

The use of a hybrid Hannonic/Stochastic model, such as the MBE one, is examined in the context of a High Quality TTS system (Fs=16 kHz). Analysis errors are studied in case of a direct analysis-synthesis scheme, and the exact responsibility of the analysis algorithm, rather than the model itself, is investigated. Through its application on well-known signals, it is found that: - Among the available analysis criteria, the Abrantes et al [2] approach slightly emerges, even though little audible improvements are obtained on real speech. - The MBE analysis of a single cosine with slowly tune-varying fundamental frequency introduces severe biases on its estimated amplitude and phase, especially for high central frequencies, while amplitude variations result in frequency-independent amplitude biases only. These effects are due to the fact that a constant frequency and amplitude is assumed during the whole analysis frame. They are responsible for the existence of HF noise in synthesized speech. Keywords : Multi-Band analysis, Analysis Criteria, Time-Varying Parameters

Full Paper

Bibliographic reference.  Dutoit, Thierry / Leich, Henri (1993): "An analysis of the performances of the MBE model when used in the context of a text-to-speech system", In EUROSPEECH'93, 531-534.