- 262 Downloads
Even though there exist various techniques to improve the recognition rates of SI systems, state-of-the-art SD systems still yield higher recognition rates than SI ones. If provided with the same amount of training data, they can achieve an average word error rate a factor of two or three lower than the SI system [Woo99]. But to train SD systems, large amounts of speaker specific speech data are needed and it is often not feasible to collect this data. Hence the use of speaker adaptation methods is appealing for solving this problem, since they promise to achieve SD performance, but require only a small fraction of speaker-specific data.
KeywordsAdaptation Data Regression Class Address Corpus Dynamic Weight Speech Frame
Unable to display preview. Download preview PDF.