1. The sample sets the mood
Zero-shot cloning copies everything about the sample: its sound, its pace and its emotional colour. A sample read in a flat, careful voice gives careful takes however you set the sliders, while a sample with life in it gives takes with life in them. Record the sample in the mood the voice will mostly be in. For a narrator that might be warm and steady; for a hero, determined; for a villain, amused and cold.
2. Pick a style
- Balanced is the default: natural delivery, close to the sample.
- Expressive raises emotion intensity, loosens the pacing and lets the delivery vary more. Good for characters, comedy, drama and shouting.
- Calm lowers emotion intensity and steadies the delivery. Good for narration, instructions and meditation.
- Fast draft is for quick English takes; emotion intensity doesn't apply in it.
3. Fine-tune the intensity
Open Fine-tune on the voice. Emotion intensity runs from flat and neutral to dramatic; Balanced sits at 0.5 and Expressive at 0.8. Move it in small steps and generate a test line each time. Past about 1.2, delivery can become theatrical or unstable, which is sometimes what a scene needs and often isn't. Expressiveness controls how varied each take is: higher gives more surprising readings, lower gives steadier, more predictable ones. Higher emotion also tends to speed speech up, so lower guidance strength a little to keep the pace natural.
4. Write the emotion into the line
A voice can only perform what's on the page. Words and punctuation do a lot of the work:
| Flat on the page | Carries the feeling |
|---|---|
| I am angry that you did that. | You did what? After everything I told you? |
| That is very good news. | Wait. Really? Oh, that's wonderful! |
| I am afraid of the dark room. | Don't. Don't open it. Please. |
- Short sentences and full stops sound tense or final.
- Questions lift the voice; exclamation marks add energy. Use them sparingly.
- Leave out stage directions like
(angrily): they'd be read aloud. - Capital letters don't add stress: the voice reads in lower case. Put the important word last instead.
5. A voice per mood, for big swings
One sample can't do everything. A character who is calm in most scenes but screams in one will sound more
convincing if you record a second sample of them shouting and add it as its own voice, for example
Mara and Mara shouting. Any line can be read by any voice in the cast, so give the
shouted lines to the second voice. Pasted lines that start Mara shouting: go to it automatically.
6. Generate, compare, keep
Emotional reads vary more from take to take than neutral ones. Generate a line two or three times, compare each new take with the one you have, and keep the best. See picking the best take.
What AI voices still find hard
- Sounds that aren't words, such as laughter, sobbing, gasps and screams. Record them yourself, or use a sound library, and place them in the edit.
- Whispering from a normal-voiced sample. Record a whispered sample as its own voice instead.
- Sudden changes inside one line, from calm to furious mid-sentence. Split the line where the mood changes.