INTERSPEECH 2004 - ICSLP
8th International Conference on Spoken Language Processing

Jeju Island, Korea
October 4-8, 2004

Prediction of the Glottal LF Parameters Using Regression Trees

Michelle Tooher, John G. McKenna

Dublin City University, Ireland

The behaviour of the glottal Liljencrants-Fant (LF) parameters were studied across vowel, context, duration, stress and fundamental frequency. Using statistical analysis we attempted to account for the variation in the glottal parameters for a speaker and from the patterns observed create a model capable of predicting the parameters for a given utterance and fundamental frequency. The parameters varied, as expected, with fundamental frequency, and both prosodic and contextual information were statistically significant predictors. Regression trees were used to create models for each of the parameters and predictions of the LF parameters were calculated. The average relative percentage error of the predictions varied across parameters, and the study indicated that it is possible to make acceptable predictions of the LF parameters.

Full Paper

Bibliographic reference.  Tooher, Michelle / McKenna, John G. (2004): "Prediction of the glottal LF parameters using regression trees", In INTERSPEECH-2004, 773-776.