Sixth European Conference on Speech Communication and Technology

Budapest, Hungary
September 5-9, 1999

Acoustical Modelling of Phone Transitions: Biphones and Diphones - What are the Differences?

S. Dobrisek, F. Mihelic, N. Pavesic

University of Ljubljana, Faculty of Electrical Engineering, Ljubljana, Slovenia

The paper presents our experiences with the phonetransition acoustical models. The phone transitionmodels were compared to the traditional context de-pendent phone models. We put special attentionon the speech signal segmentation analysis to pro-vide a better insight into certain segmentation effectswhen using the different acoustical models. Experi-ments with the HMM-based models were performedusing the HTK toolkit, which was extended to allowproper state parameter tying for the phone transitionmodels. All the model parameters were estimatedon the GOPOLIS speech database. The annotationconfusions concerning two-phone speech units are dis-cussed. Finally, the overall word recognition score ispresented. The better score was achieved using thediphone models even when comparing them to thetriphone models.

Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Dobrisek, S. / Mihelic, F. / Pavesic, N. (1999): "Acoustical modelling of phone transitions: biphones and diphones - what are the differences?", In EUROSPEECH'99, 1307-1310.