Vits fine-tuning question

[gundam233]

Hey folks! Quick question about fine-tuning the LJSpeech-VITS model.
I trained it for 20k+ steps with default configs, but the output audio still doesn’t sound good (issues with timbre and natural pauses). I watched some tutorials that said I should freeze the encoder once the voice’s timbre/pauses feel decent, then keep training.

Wondering if anyone can confirm:

• Is freezing the encoder a necessary step here?
  ​
• What’s the correct workflow for this (e.g., when to freeze, what to adjust after freezing)?
  ​
• Any tips to avoid wasting steps and get better results faster?