how to impove the quality of GLA when T2 model is trained by setting magnitude_power=2

Rayhane-mamah / Tacotron-2

DeepMind's Tacotron-2 Tensorflow implementation

MIT License

2.28k stars 905 forks source link

how to impove the quality of GLA when T2 model is trained by setting magnitude_power=2 #386

Open NewEricWang opened 5 years ago

NewEricWang commented 5 years ago

Hi, I found the quality of speech generated by GLA descend when training model based on magnitude_power=2. The spectrum of generated speech is discontinuous, especially at off-frequency. How to improve the quality for GLA? Which parameters can be adjusted?