Currently voice customizability is an experimental feature. Over the next couple weeks, we plan to improve it substantially and make it more reliable. Right now, some known issues include:
  • Slider behaviors do not exactly match their labels (e.g. adjusting pitch slider doesn't seem to change only pitch)
  • Adjusting sliders can introduce quality issues like graininess
  • During conversation, voice unexpectedly switching to other base voices, tone, and/or gender
And other problems with consistency and reliability.