Turns a written scene into acted, multi-voice speech and redoes a take when you give it a note

You write the words and the acting in the same box. Short directions go in square brackets where they should happen, like [sighs], [laughs] or [pause]. A stretch wrapped in <whisper> and </whisper> is whispered. A mood at the start, like [icy calm], sets the tone for the whole line. None of the bracketed words are read out; they only tell the voice how to perform. Then you pick who speaks: clone a voice from a short clip, describe one in plain words ("gravelly old sea captain, slow, salt in every word"), or let the model invent one.

What is new is the retake. If a line comes out too gloomy, you write a note such as "more amused than before; drop the hurt" and it performs the same words again, judging the note against the take you just heard.

The same model handles whole scenes. One script can hold up to 16 characters, each line labeled with who is speaking, and long text is split into chunks and streamed as it is generated. It can also correct a finished line: give it the old take and the fixed sentence, and it reads the change in the same voice.

Fish Audio's official Space runs the full model on a dedicated GPU, so you can direct it in your browser right now. On the Studio tab, click the "Bad news" example and press Perform: you hear the result read with heavy sadness. Now type "more amused than before; drop the hurt" in the note box and press Retake. The same sentence comes back with a smile in it. Then switch the voice to Design and pick the Sea captain to hear a character built from a description alone. Guests get 5 takes a day, and signing in with a free Hugging Face account adds 20 more. Running it yourself needs a data-center NVIDIA card and the SGLang Omni server.

Takes sometimes stall: the model can fall silent mid-line or fail to finish, so a long line may need a second try. The license is Fish Audio's own. Personal and research use is free, and commercial use is allowed while your organization earns under $500,000 a year, but offering it to others as a public voice service needs a separate deal. Only clone voices you have permission to use.

• Voice a game cutscene or short animation from one script, with a different voice for every character.
• Record an audiobook chapter where the dialogue sounds acted instead of read aloud.
• Try three readings of an ad line, each with a different note, and keep the one that lands.
• Fix a wrong date in a generated voiceover without redoing the whole take.
• Sketch a radio drama or podcast skit before you book real actors.

Try the demo

View model page

Read this issue on neuralexpedition.com →