14th International Congress of Phonetic Sciences (ICPhS-14) San Francisco, CA, USA August 1-7, 1999

Neural Network Models for Finnish Prosody

Martti Vainio (1), Toomas Altosaar (2), Matti Karjalainen (2), Reijo Aulanko (1), Stefan Werner (3)

(1) University of Helsinki, Department of Phonetics, Finland
(2) Helsinki University of Technology, Acoustics Laboratory, Finland
(3) University of Joensuu, Department of German, Finland

Ongoing research on Finnish prosody with respect to modeling pitch, loudness and segmental duration with artificial neural networks is presented. The neural network methodology is a continuation of our earlier studies on word level prosody. We investigated the problem of generating values for prosodic features from symbolic input with standard multi-layer feed-forward networks that were trained with back propagation. Our results show that the neural network methodology is well suited to modeling and studying all phonetic aspects as well as levels of prosody.

Full Paper

Bibliographic reference.  Vainio, Martti / Altosaar, Toomas / Karjalainen, Matti / Aulanko, Reijo / Werner, Stefan (1999): "Neural network models for Finnish prosody", In ICPhS-14, 2347-2350.