Third ESCA/COCOSDA Workshop on Speech Synthesis

November 26-29, 1998
Jenolan Caves House, Blue Mountains, NSW, Australia

Parametric Modeling of Intonation using Vector Quantization

Gregor Möhler (1), Alistair Conkie (2)

(1) Institute of Natural Language Processing, University of Stuttgart, Germany
(2) AT&T Labs - Research, Florham Park, NJ, USA

In this study we propose a data-based approach to intonation modeling using vector quantization. The model is based on an F0 parametrization with an especially designed approximation function. The parameter vectors found are vector quantized with varying codebook sizes. This method is motivated by intonation theories that suggest that pitch accent and boundary phenomena can be described by a distinct number of di erent types. We use classification trees to predict the F0 movements represented in the codebook from a set of features. We assessed the quality of the model by numerical measures and perceptual testing. The tests show that our method performs well when compared with other methods of intonation modeling.


Full Paper

Bibliographic reference.  Möhler, Gregor / Conkie, Alistair (1998): "Parametric Modeling of Intonation using Vector Quantization", In SSW3-1998, 311-316.