Audio in, audio out
Speech-to-speech
A single realtime model takes the caller’s audio and answers in audio. There is no transcription or voice step in between, so there is less to wait for.
- Realtime models
- 7 models
- Providers
- OpenAI and xAI
- Voices
- 17 voices
- Listening
- Thinking
- Speaking
The same question, two engines
Switch between engines to see what disappears when the model works in audio.
Talk to a speech-to-speech agent
This agent runs on a realtime speech model. Listen for how quickly it starts answering.
Seven realtime models, one setting
Choose the realtime model per agent. Each brings its own voices and its own way of deciding when you are done talking.
OpenAI realtime models
Semantic voice activity detection judges whether the caller finished a thought, not just whether they went quiet.
- GPT Realtime
- Realtime Mini
- Realtime 2
- Realtime 2.1
- Realtime 2.1 Mini
- GPT-Live-1
- 12 voices
xAI Grok Voice Agent
xAI’s realtime voice model, with turn detection handled on the server.
- 5 voices
- 40+ languages
- Server-side VAD
Which engine fits?
Same tools, knowledge and phone numbers on all three. The engine is chosen per agent.
- How it works
- Speech-to-text, then a language model, then text-to-speech
- Response feel
- Fast, with every stage streaming
- Control
- Highest. Swap STT, model and voice independently
- What you can inspect
- Transcript, model and voice latency, end to end
- Reach for it when
- You need a specific voice, a specific model, or strict tool use
- How it works
- One realtime model takes audio in and returns audio
- Response feel
- Fastest to first word; no transcription hop
- Control
- Choose the model and its voice; one provider per call
- What you can inspect
- Transcript and end-to-end latency
- Reach for it when
- Speed and natural tone matter more than per-stage control
- How it works
- Keeps listening while it speaks, so callers can overlap and interrupt
- Response feel
- Most like a person; overlaps, backchannels, barge-in
- Control
- Tune interruption sensitivity and listening cues
- What you can inspect
- Full transcript and call detail
- Reach for it when
- Callers talk over each other and the agent must keep up
| Attribute | CascadingSTT → LLM → TTS | Speech-to-speechAudio in, audio out | Full duplexListens while it talks |
|---|---|---|---|
| How it works | Speech-to-text, then a language model, then text-to-speech | One realtime model takes audio in and returns audio | Keeps listening while it speaks, so callers can overlap and interrupt |
| Response feel | Fast, with every stage streaming | Fastest to first word; no transcription hop | Most like a person; overlaps, backchannels, barge-in |
| Control | Highest. Swap STT, model and voice independently | Choose the model and its voice; one provider per call | Tune interruption sensitivity and listening cues |
| What you can inspect | Transcript, model and voice latency, end to end | Transcript and end-to-end latency | Full transcript and call detail |
| Reach for it when | You need a specific voice, a specific model, or strict tool use | Speed and natural tone matter more than per-stage control | Callers talk over each other and the agent must keep up |
Tell us what you are building
Tell us what you are making. We will walk you through the engines, the API, and what it takes to run it in production.