Iterative English accent adaptation in a speech synthesis system

The paper documents some initial results of our investigation of the applicability of adaptation techniques and procedures for use in speech synthesis systems - a process whereby a well-trained synthesizer can be transformed to sound "more like" a small target recording set. If we can ensure that the target data set is representative not of an individual speaker, but of a category of accented speech, adaptation of the synthesis system could be expressed as an iterative procedure where the system learns to modify both its spoken output and linguistic representation to resemble more closely the alternative accent. We summarize the proposed algorithm and present results estimating the rate of pronunciation-learning and the generalization possible with this technique, specifically as regards the choice of an appropriately-sized target speech database for training.