Voices
Loading…
Try a voice
Temperature
1. Reference clips
12 to 20 seconds of one person speaking calmly in a quiet room, in one to three clips, each with exactly what is said. WAV or FLAC is best; any format the browser plays works. See the guide.
The TTS service reads a script in a voice of its own: a voice no real person owns, with a transcript known word for word. Each candidate is different. Generate a few, keep the one you like that passes its checks. About a minute each on an RTX 3090.
Temperature
2. Make the voice
Add a clip to see whether the set is enough.
One voice per delivery: a warm and a brisk take of the same speaker are two voices (ginger-warm, ginger-brisk). Pick one per broadcast.