Docs
Voice and speed
Ultra mode
Get faster, more natural turn-taking. One speech-to-speech model listens and answers directly.
In standard mode a call passes through three models: speech-to-text, a language model and a voice. In Ultra mode a single full-duplex model listens and speaks at once. There are no hand-offs between models, so replies start sooner and the timing of the conversation feels more human.
| Standard | Ultra | |
|---|---|---|
| Pipeline | Speech-to-text, then a model, then a voice | One speech-to-speech model |
| Voices | The full library, plus cloned voices | Marin, Cedar, Alloy, Ash, Ballad, Coral, Echo, Sage, Shimmer and Verse |
| Tuning | Every part can be changed separately | Fewer settings |
Instructions, tools, knowledge and analysis work the same way in both modes.