accuracy compares with VITS?

[lucasjinreal]

accuracy compares with VITS? does it faster and accurator?

[RayeRen]

We didn't compare with VITS in our paper quantitatively. Perceptually, the sound quality and prosody of PortaSpeech and DiffSpeech are similar with VITS. The differences between our models and VITS are:

• Our models are acoustic models, which need an external vocoder (HiFi-GAN in our case) to build a TTS system, while VITS is an end-to-end TTS model.
• Our models can train faster than the end-to-end TTS model since we can fix the vocoder part, which makes the model tuning easier.
• End-to-end TTS models do not need to consider the quality gap between the generated and ground-truth intermediate features (mel-spectrogram), which avoids the error propagation problem.

So e2e TTS model and non-e2e TTS model have different usage scenarios.

[MaxMax2016]

i think PortaSpeech + MB MelGAN can work on cpu very well, VITS need gpu to work.

[RayeRen]

i think PortaSpeech + MB MelGAN can work on cpu very well, VITS need gpu to work.

Yes, you can use other vocoders with PortaSpeech.