Figure 3.
The chart shows horizontal box plots with the x-axis labelled naturalness M O S ranging from 1 to 5. Categories include G T, Tacotron 2 clean, Tacotron 2 W E R 070, Tacotron 2 W E R 358, V I T S clean, V I T S W E R 070, V I T S W E R 358, Grad T T S clean, Grad T T S W E R 070, Grad T T S W E R 358, D V T clean, D V T W E R 070, D V T W E R 358, D V T 2 clean, D V T 2 W E R 070, and D V T 2 W E R 358. Each shows distribution ranges with central markers around values between 3 and 5, with G T highest and degraded W E R conditions lower.Mean opinion score (MOS) results for naturalness evaluation, with 95 % confidence intervals. The results show that diffusion-based models (GradTTS, DVT, DVT2) maintain higher naturalness when trained on noisy data compared to VITS and tacotron 2. Previously presented inFeng et al. (2024)