8th European Conference on Speech Communication and Technology

Geneva, Switzerland
September 1-4, 2003


Experimental Evaluation of the Relevance of Prosodic Features in Spanish Using Machine Learning Techniques

David Escudero (1), Valentin Cardenoso (1), Antonio Bonafonte (2)

(1) Universidad de Valladolid, Spain
(2) Universitat Politecnica de Catalunya, Spain

In this work, machine learning techniques have been applied for the assessment of the relevance of several prosodic features in TTS for Spanish. Using a two step correspondence between sets of prosodic features and intonation parameters, the influence of the number of different intonation patterns and the number and order of prosodic features is evaluated. The output of the trained classifiers is proposed as a labelling mechanism of intonation units which can be used to synthesize high quality pitch contours. The input output correspondence of the classifier also provides a bundle of relevant prosodic knowledge.

Full Paper

Bibliographic reference.  Escudero, David / Cardenoso, Valentin / Bonafonte, Antonio (2003): "Experimental evaluation of the relevance of prosodic features in Spanish using machine learning techniques", In EUROSPEECH-2003, 2309-2312.