Sixth International Conference on Spoken Language Processing
(ICSLP 2000)

Beijing, China
October 16-20, 2000

Data Driven Intonation Modelling of 6 Languages

Jeska Buhmann (1), Halewijn Vereecken (1), Justin Fackrell (2), Jean-Pierre Martens (1), Bert van Coile (1,2)

(1) ELIS, Ghent University, Ghent, Belgium
(2) Lernout & Hauspie Speech Products NV, Ieper, Belgium

A method for creating multi-lingual intonation models is described. The method adheres closely to the pioneering work of Traber, in that a recurrent neural network (RNN) predicts a number of F0 values per syllable. An important aspect of the work presented here is the selection of linguistic and prosodic features that are suitable for predicting the observed intonation phenomena in different languages. Another aspect is the use of automatic labelling techniques for the preparation of the training data. Experiments on six languages demonstrate that even though there are differences in performance across languages, it is possible to obtain good results for all six languages. More importantly, making use of automatic labelling techniques for the construction of the training corpora, tends to give better results than making use of manual labelling techniques.

Full Paper

Bibliographic reference.  Buhmann, Jeska / Vereecken, Halewijn / Fackrell, Justin / Martens, Jean-Pierre / Coile, Bert van (2000): "Data driven intonation modelling of 6 languages", In ICSLP-2000, vol.3, 179-182.