In an effort to positively influence the area of automatic speaker characterisation, we draw on existent technologies to efficiently derive the voice source and subsequently learn its behaviour in a varierty of environments.

We observe the behaviour of the Liljencrants-Fant (LF) voice source parameters as fitted to closed-phase inverse filtered vowel data from a corpus of two speakers, male and female. The LF parameters are tracked as they vary across vowel, phonetic environment, duration, power and fundemental frequency. Statistical analysis is applied to the data in an effort to pinpoint those phenomena that attribute to variation in parameters within and across speakers.

Non-linear regression trees and backpropagation neural networks are built for each parameter and tested for learning capability.

The learnt knowledge is applied to synthesis and recognition experiments in order to guage the performance of such speech systems when speaker specific information is included. We believe that this knowledge of voice source variation both within and among speakers has the potential to influence not only speech recognition, but many areas of speech technology

Tooher, M., McKenna, J., (2003) "Variation of the glottal LF parameters across F0, vowels, and phonetic environment", Proc. VOQUAL ISCA Workshop, Geneva 2003.

Tooher, M., McKenna, J., (2004) "Prediction of the Glottal LF Parameters using Regression Trees", to appear in Proc. Interspeech 2004, Korea.