What survives the swap
Your timing and emphasis carry over, because the performance is still yours
Word for word: nothing is re-transcribed and nothing is read again
A large set of target voices, including the ones you have cloned yourself
Keeps takes consistent when the original speaker is no longer available
Works on narration, on dialogue and on a scratch track alike
Your upload is processed to produce the result and never used to train a model
Use cases
Narration recorded as a placeholder that now wants a finished voiceDialogue for characters, acted once and voiced several waysKeeping a series consistent after the original reader has moved onProtecting a contributor without flattening what they saidTrying a script in several voices before committing to oneReading a line yourself because the timing matters, then swapping the voice
Changing a voice, answered
How is this different from text to speech?
Text to speech reads words off a page and decides the delivery itself. This takes a recording you already made and keeps the delivery, swapping only the voice, so the performance is still yours.
Is this the same as cloning a voice?
No, and the two work together. Cloning builds a new voice out of a sample; changing a voice applies a voice, cloned or ready-made, to a recording that already exists.
What audio should I upload?
One speaker, little background noise and no music underneath. Crosstalk and a noisy room cost accuracy here for the same reason they do in transcription.
Can I use a voice I cloned?
Yes. A voice you have cloned appears as a target alongside the ready-made ones, which is the usual way to keep a whole project in one voice.