Enter the app password to continue. (Demo: any value works.)
Defaults (0.5 / 0.5) suit most voices. For more expressive speech try exaggeration ≈ 0.7 with CFG ≈ 0.3. If the speaker is fast, lower CFG to ≈ 0.3. To keep a regional accent (e.g. Quebec French), use Faithful accent: a higher CFG follows the reference more closely. French voices start on it until you change a setting. A fixed seed makes results reproducible.
Nothing generated yet.
Training takes the GPU: generation pauses until the run ends (3–6 h on a 24 GB card for the Quebec recipe). The result appears in the model list once validated.
No training run yet.