Fast, spotted text-to-speech model with voice cloning
Pick a bundled voice, or clone one from your own recording.
Sampling temperature; lower = more stable, higher = more varied.
Keep only the k most likely tokens per head (0 = disabled).
Cap on generated frames (~21.5 fps; 43 frames ≈ 2 s, 1075 ≈ 50 s max).
Penalize recently used tokens (1.0 = disabled).
Recent frames tracked by the penalty (0 = full history).