| 摘要 |
PURPOSE:To always obtain a high speech recognition rate irrelevantly to the number of learning samples used for speaker adaption. CONSTITUTION:A model is learnt by three kind of speaker adapting methods, e.g. speaker connection mixed weight learning method(STWT)1, a speaker connectionless mixed weight learning method(SFWT) 2, and a moving vector field smoothing system(VFS) 3 to generate adaption models 4, 5, and 6 respectively. Then, speech recognition is carried out by the respective adaption models 4, 5, and 6 and the result having maximum likelihood is selected and outputted as a recognition result. |
