Intonation: Theory, Models, and Applications

Athens, Greece
September 18-20, 1997


Coding of Large Intonational Units by Linear Prediction

Joachim Mersdorf, Ansgar Rinscheid, Marc Brüggen, Kai Uwe Schmidt

Institute of Communication Acoustics, Ruhr-University, Bochum, Germany

A new approach for analyzing, coding, predicting and generating F0 curves based on linear predictive coding [1] is presented. The suggested model supports an automatic extraction of different parameters and accents from intonational contours by means of an appropriately simplified rectangular residual function and typical filter coefficients. The signal approximation can be regarded as an excitation signal for prosodical accents and other gestures, as a result of the pragmatical, semantical, syntactical and syllabical structure of the sentence. The resulting filter coefficients can be interpreted either as physiological parameters [3][4] or as speaker-individual features in the intonation shape. The method was developed for both practical prosodical data analysis and contour resynthesis in TTS or dialog machines. Perceptual evaluation also indicated first success for invariant filter structures and highly simplified residual signals. In this first approach the model is perceptually motivated [5] and directed towards the intonational gestalt on the sentence or phrase level.

Full Paper

Bibliographic reference.  Mersdorf, Joachim / Rinscheid, Ansgar / Brüggen, Marc / Schmidt, Kai Uwe (1997): "Coding of large intonational units by linear prediction", In INT-1997, 235-238.