Why not use the final mel output as the prenet input?

Rayhane-mamah / Tacotron-2

DeepMind's Tacotron-2 Tensorflow implementation

MIT License

2.28k stars 905 forks source link

Why not use the final mel output as the prenet input? #429

Open terryyizhong opened 5 years ago

terryyizhong commented 5 years ago

I have a question about the inference architecture of Tacotron2. I know the paper use the mel of decoder output as the prenet input. But Why not use the final mel output (decoder output combined with postnet output, which I think is closer to the ground truth mel) as the prenet input?