Latest MOS study results for TTS

Based on our latest MOS study, TTS is able to outperform any other solution given a clear dataset.
Model details:

Judy Wave1: Tacotron + WaveRNN
Judy Wave2: Tacotron2 + WaveRNN
Judy GL1 : Tacotron + Griffin Lim
Judy GL2 : Tacotron2 + Griffin Lim

1 Like

What is the “Judy” dataset? Is it publicly available?

Great result, by the way. Interesting that Taco1 outperforms Taco2 in “Good” category but has more “poor” results. Otherwise, they look on par.

Is “Judy” Taco1 pretrained model available?

Judy is an internal dataset 25 hours long. The only difference is that it is recoded in better condition than LJSpeech.

Is the dataset created by Mozilla? Any chance it could be shared with the community?

I don’t think that would be possible in near future. Not because we don’t like but it is just too expensive with the current hype and its recognition by the voice talents.