Last updated
What is OpenAI Realtime used for in a Bolna voice agent?
OpenAI Realtime lets a Bolna agent run on one speech-to-speech model that takes in caller audio and speaks the reply, instead of chaining a transcriber, an LLM and a synthesizer. Because speech is never turned into text and back, the model hears how something was said, and the reply arrives over one connection rather than three. It handles turn taking, calls your tools and keeps a transcript of both sides.
How do I set up OpenAI Realtime with Bolna?
- In Agent Studio, switch the agent's Voice engine from Cascaded to Realtime (Beta).
- Use the Model tab, which replaces the Intelligence and Languages tabs, to pick the model and voice.
- In the API, set
agent_typetorealtime_agent, addtools_config.s2swithproviderset toopenai_realtime, and set the toolchain pipeline to[["s2s"]]. Leave the transcriber, LLM and synthesizernull. - Choose a model, such as
gpt-realtime-2.1, and a voice, such asmarin. - Set
languageto the caller's ISO code. Without it, the model detects language per utterance and can transcribe into the wrong script. - Keep
turn_detection_typeatsemantic_vadand tuneeagernessif the agent replies too early or too late.
Every field is listed in the OpenAI Realtime setup guide in the Bolna docs.
Which OpenAI Realtime models does Bolna support?
| Model | Notes |
|---|---|
gpt-realtime-2.1 | Default. Latest realtime model with reasoning and tool use |
gpt-realtime-2.1-mini | Cost-efficient version of gpt-realtime-2.1 |
gpt-realtime-2 | Previous generation with reasoning and tool use |
gpt-realtime-1.5 | Audio in, audio out, without reasoning |
Voices: alloy, ash, ballad, cedar, coral, echo, marin, sage, shimmer and verse. gpt-realtime-translate is not supported for agents, and gpt-realtime-mini is not offered.
What can you build with OpenAI Realtime and Bolna?
Agents where responsiveness matters most. Semantic VAD judges whether the caller has finished a thought instead of waiting on a silence timer, so the agent interrupts less and replies sooner. Raise eagerness for faster turns or lower it for callers who pause.
Tool-using agents. Reasoning models call tools during the call, and reasoning_effort runs from minimal to xhigh when an answer needs more thought. Custom tools with pre-call webhooks, end_call and transfer_call carry over from standard agents.
When to keep the classic pipeline. Speech-to-speech does not support knowledge bases, graph agents or multilingual configuration. If you need those, or a specific transcriber or voice, use OpenAI GPT with a separate transcriber and synthesizer.
Comparing realtime providers. Gemini Live is the other speech-to-speech option on Bolna, with 30 prebuilt voices and session resumption for long calls.