Why Paraphrase never generates speech
Bottom line: Paraphrase cannot put a word in your mouth, because an edit in Paraphrase is a list of positions in your own transcript and there is no audio for a word you never said. That is a deliberate choice. If you need to fix a word you did not record, a tool with an AI voice will do it and Paraphrase will not.
Many transcript-based editors now include a generated voice: type a word that is not in the recording and the tool synthesises it in your voice. It is clever and it is useful. Paraphrase does not do it, and it never will.
What an edit is in Paraphrase
When Paraphrase transcribes a take, every word comes back with the slice of audio that produced it. Every edit you make, and every edit the cleanup proposes, is a statement about which of those words survive and in what order. When you export, the original recording is cut at those words and the pieces are joined back up.
So the set of things an edit can express is exactly the set of things you said. Removing a word is expressible. Moving a word is expressible. A new word is not, because nothing in that list points at audio that does not exist.
Why that matters
The export is evidence of what you said. A podcast guest, an interviewee or a lecturer can be cut, tightened and reordered, but not made to say something new. If you publish interviews, that is a property worth having.
The AI cannot go further than you would. The cleanup is a language model, and language models can be confidently wrong. In Paraphrase the worst it can do is propose cutting or moving a word you wanted left alone, and every cut it proposes shows struck through for you to restore. It cannot invent a phrase to smooth over a cut.
There is no voice model of you. Nothing is trained on your recordings, by us or anyone else, and there is no clone of your voice sitting on a server.
What you give up
Being honest about this: sometimes a generated word is exactly what you want. You misspoke a name, the take is otherwise perfect, and re-recording is not an option. A tool with an AI voice can fix that. Paraphrase cannot. Your options here are to cut around it, or record a pickup take into the same project and use the correct word from that.
If that trade does not suit you, Descript can. Its Regenerate tool lets you highlight script, type new words and have them generated in a custom AI speaker, and its custom voice clones need recorded authorisation from the person whose voice is cloned. Our Descript comparison covers the difference in detail.
Try it
A new free account comes with fifteen minutes of transcription and three AI cleanups a day, no card needed. Start on paraphrase.bot and see what the cleanup proposes for your own recording.