Intonation: Theory, Models, and Applications

Athens, Greece
September 18-20, 1997


Modeling Finnish Microprosody for Speech Synthesis

M. Vainio (1), Toomas Altosaar (2), Marti Karjalainen (2), R. Aulanko (1)

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

In our research on the prosody of Finnish we have modeled word-level prosody with artificial neural networks and a comparable model based on multiple linear regression. In earlier work [6] we have used multi-layer feed-forward networks to estimate prosodic parameters - segmental durations, loudness and pitch values - for single phones within words. The performance of these models in predicting average values for phones was encouraging. Nevertheless, a more detailed account of the behavior of the continuous prosodic parameters is needed, e.g., to enhance the intelligibility and naturalness of speech synthesis. We present results for both models' ability to estimate microprosodic variation of pitch in Finnish.

Full Paper

Bibliographic reference.  Vainio, M. / Altosaar, Toomas / Karjalainen, Marti / Aulanko, R. (1997): "Modeling Finnish microprosody for speech synthesis", In INT-1997, 309-312.