Figure 6.
The five bar charts are labelled Tacotron 2, V I T S D P, Grad T T S, D V T, and D V T 2. The vertical axis shows W E R percent. Each chart includes five categories N style, 2 x length, H connect, T connect, and H T connect, with a baseline value shown. Tacotron 2 shows 24.4, 25.1, 21.2, 20.1, and 20.8 with a baseline of 24.4. V I T S D P shows 19.6, 19.8, 10.1, 13.9, and 17.2 with baseline 19.6. Grad T T S shows 17.5, 18.0, 9.4, 5.9, and 6.9 with baseline 17.5. D V T shows 18.0, 18.4, 7.3, 5.2, and 4.9 with baseline 18.0. D V T 2 shows 18.8, 18.9, 7.1, 5.2, and 4.8 with baseline 18.8.WER Results of synthesized speech by WER-358 noisy models of all systems inferred on the test set, and the WER results after using the “text-prompt” connection. We directly concatenate “text-prompt” and “synthesized-target” as inputs for inference. For example, using a HT-connect, if “text-prompt” is “good morning” and “synthesized-target” is “have a nice day,” the concatenated text would be “good morning have a nice day good morning”