Skip to content

Mean Opinion Score Study

Eren Gölge edited this page Jan 12, 2021 · 1 revision

Based on our latest MOS study, TTS is able to achieve on par performance with any other solution. However, it is also shown that the quality of the dataset is an important part of TTS. Therefore, it is worth to find a good resource or even record your own dataset, if you like to peak the TTS performance.

You can find some pointers under Dataset

Models

  • Judy Wave1: Tacotron + WaveRNN
  • Judy Wave2: Tacotron2 + WaveRNN
  • Judy GL1 : Tacotron + Griffin Lim
  • Judy GL2 : Tacotron2 + Griffin Lim
  • MozillaTTS Nancy: Tacotron + Griffin Lim
  • MozillaTTS Nancy2: Tacotron2 + WaveRNN
  • MozillaTTS LJSpeech: Tacotron + GriffinLim

(Judy is an 25 hours long internal dataset)

(.Jofish .Abe and .Janice are real human voices)

Clone this wiki locally