Intonation: Theory, Models, and Applications

Athens, Greece
September 18-20, 1997


Intonation Estimation for Romanian Language

I. Gavat, M. Zirra, O. Cula

"Politehnica" University of Bucharest, Department of Electronics and Telecommunications, Bucharest, Romania

We propose a method to estimate prosodic features from pitch contours combined with energy contours, extracting information about: the lexical stress, the melody in the sentence and the duration of all segments in the utterances.

To determine F0 contours, we have utilised two methods for pitch detection: a nonlinear variant for a cepstrum pitch detector which works very effectively in the range of medium pitch periods and stationary segments and an event based pitch detector applying the dyadic wavelet transform proposed by Mallat, which can correctly estimate the pitch variation for nonstationary and also for noisy segments of large varying length. The speech material consisting in 50 sentences, 36 affirmative, 7 exclamative and 7 interrogative, uttered by 3 male and 2 female speakers, was analysed by combining pitch and energy contours, and the results were compared with results obtained by manually segmenting and stressing of the same speech material. In mean 92,59% of the situations were correct for lexical stress, 90,3% for word boundaries and 84% for global phrase intonation.

Full Paper

Bibliographic reference.  Gavat, I. / Zirra, M. / Cula, O. (1997): "Intonation estimation for romanian language", In INT-1997, 149-152.