14th International Congress of Phonetic Sciences (ICPhS-14)

San Francisco, CA, USA
August 1-7, 1999


A Classification of Polish Pitch Accents Using Neural Networks.

Grażyna Demenko (1), Wiktor Jassem (2)

(1) A. Mickiewicz University in Poznan, Institute of Linguistics, Poland
(2) Polish Academy of Sciences, Institute of Fundamental Technological Research, Poznafi

Neural networks were used to classify nuclear and prenuclear tunes. Experiment (1) included imitations of one- or two-syllable phrases representing 9 different nuclear tunes defined in terms of Low, Mid or High tone. Exp. (2) included imitations of longer phrases, some with an unaccented PTs, H or L PTs and the 9 NTs of exp. (1).
   All imitations were judged for equivalence with the models, and wrong imitations were discarded. In (1) and (2) one part of the materials was used for training and another part for testing. Exp. (3) included a read text, in which some of the classes of tunes distinguished in exp. (1) and (2) were conflated so as to produce 4 classes of nuclear tunes. The accuracy of classification in the training set and the testing set was similar: The scores were mostly contained within a bracket of approx. 80 to 90 percent.

Full Paper

Bibliographic reference.  Demenko, Grażyna / Jassem, Wiktor (1999): "A classification of Polish pitch accents using neural networks.", In ICPhS-14, 1525-1528.