Table 1 Algorithm for initializing parameters
From: Non-parallel dictionary learning for voice conversion using non-negative Tucker decomposition
Initializing in the parallel setting |
∙ Set source and target parallel data Xs and Xt |
∙ Optimize Ws, Wt, and H minimizing dKL(Xs,WsH)+dKL(Xt,WtH) |
∙ Optimize Us, Ut, and G minimizing dKL(Ws,UsG)+dKL(Wt,UtG) |
Initializing in the non-parallel setting |
∙ Set source training data Xs |
∙ Optimize Ws and Hs while minimizing dKL(Xs,WsHs) |
∙ Set target training data Xt |
∙ Optimize A and Ht while minimizing dKL(Xt,AWsHt) while fixing Ws |
∙ Optimize Us, Ut, and G while minimizing dKL(Ws,UsG)+dKL(AWs,UtG) |