Archives of Acoustics, 23, 2, pp. 159-178, 1998
Acoustic classification and automatic border of phrase and stress recognition in Polish
A classification of perceptively and linguistically verified intonation structures was made basing on acoustic qualities of speech signals. The importance of the eight selected parameters describing the set of elements under investigation was tested by means of statistical methods. Findings of discriminant analysis proved the possibility of a correct classification of unstressed vowels (in the range of 86-92%). For the remaining vowels (stressed and located at the end of a phrase) a significant degree of correct classification was obtained in the range of 54-73%. Far better classification results (88% at the average) were obtained by using a neural network. Identification of accents and phrase boundaries originating from the new data set showed the necessity to extend the linguistic base and repeat the network training.
Full Text: PDF
Copyright © Polish Academy of Sciences & Institute of Fundamental Technological Research (IPPT PAN).