Knowledge-rich speaker adaptation for speech recognition
The objective of the project is research for increasing the performance of automatic speech recognition, especially for atypical speaker characteristics and speaking styles. A new approach will be developed for combining knowledge-based and data-driven techniques. Existing knowledge on acoustic phonetics, speech production, speaker characteristics and speaking styles will be explored in order to determine the realization of important speaker dimensions in the acoustic domain. These relations will be used to transform and extend conventionally trained models to reflect prperties which did not exist in the training data. In this fashion, the recognition accuracy will be more effective since the new models are constrained to be coherent with the used knowledge. The technique will reduce the required size of the training data considerably compared to conventional training. The results will also have practical importance for data-driven speech synthesis, pronunciation evaluation in second language learning and speaker verification. International co-operation is planned in a network of Nordic universities and American research institutes.
Group: Speech Communication and Technology
Funding: VR (2006-4313)
Duration: 2007-01-01 - 2009-12-31
Keywords: speech recognition, speaker adaptation, speaker characterization
Proceedings of ISCA ITRW Speech Analysis and Processing for Knowledge Discovery. Aalborg, Denmark. [abstract] [pdf] (2008). Investigating Explicit Model Transformations for Speaker Normalization. In