The four styles
| Style | Emotion intensity | Guidance strength | Expressiveness | Use it for |
|---|---|---|---|---|
| Balanced | 0.5 | 0.5 | 0.8 | The default, and the place to start. |
| Expressive | 0.8 | 0.3 | 0.9 | Characters, comedy, drama, energy. |
| Calm | 0.3 | 0.6 | 0.6 | Narration, instructions, meditation. |
| Fast draft | A quicker model (Turbo), English only, needing a clip of at least 5 seconds | Quick drafts and checking a script's flow. | ||
Picking a style moves the sliders. Move any slider yourself and the style shows as Custom. Whatever the sliders are set to is saved with the voice the next time you generate, so each voice comes back the way you left it.
The sliders, one by one
Emotion intensity (0 to 2)
From flat and neutral to dramatic. Higher values push the delivery further from a plain reading: more energy, bigger swings, more emphasis. Small steps make a real difference. Above about 1.2 delivery can turn theatrical or unstable. It also tends to speed speech up, so pair a raised emotion with slightly lower guidance strength. Not used in Fast draft.
Guidance strength (0 to 1)
How closely the take follows the voice clip, especially its pacing. Higher keeps takes close to the sample's rhythm; lower gives the model more freedom. If takes rush because the sample was quick, try about 0.3. When a voice speaks a different language from its sample, a very low setting, even 0, can reduce how much of the sample's accent carries over. Not used in Fast draft.
Expressiveness (0.05 to 1.5)
How varied each take is. Low values give steady, predictable readings that sound alike from take to take; high values give livelier, more surprising ones, and occasionally odd ones. Lower it for long narration where every line should match; raise it when you want options to choose from. This is the one main slider that also works in Fast draft.
The advanced three
- Repetition penalty (1 to 3): keep it near 1.2. It discourages the model from repeating itself, but set higher it makes the voice keep talking past the end of the line. In abcVoiceover's own tests, at 2.0 almost half of short lines ran on; at 1.2, none did.
- Min probability (0 to 0.2): higher is safer and more predictable, lower allows rarer choices. The default of 0.05 suits nearly everything. Not used in Fast draft.
- Top P (0.1 to 1): lower gives a more conservative delivery. Leave it at 1 unless takes are erratic.
Language
Each voice also has a Language: which of the 23 it speaks. It belongs to the voice, not the script, so a voice set to French reads French in every script it's in. To have the same person speak two languages in one project, add the clip twice, as two voices with different languages. See the language pages.
A simple way to tune a voice
- Start on Balanced and generate three test lines: a statement, a question and an emotional line.
- If they're too flat, try Expressive. Too much, try Calm.
- Then adjust one slider at a time, in small steps, and generate the same test lines again.
- When it sounds right, leave it alone for the rest of the project, so lines made later match lines made now.
What sliders can't fix
Settings shape delivery, but they can't remove echo or noise from a sample, change the voice's basic sound, or fix a mispronounced word. For those, see recording a voice sample, fixing mispronunciations and the troubleshooting guide.