OpenAI's Realtime API is the fastest way to a genuinely conversational agent: speech in, speech out, with interruption handled natively. The hard part was never the model — it was the telephony, the carrier and the compliance around it. Vomyra is that part.
gpt-realtime is OpenAI's production speech-to-speech model, generally available since August 2025 through the Realtime API. It streams audio to the model and streams audio back, keeping the whole exchange inside one session rather than chaining separate transcription, reasoning and synthesis services. The GA release added remote MCP server support, image input and native SIP phone calling, alongside two voices built for it — Cedar and Marin.
Getting GPT Realtime working in a browser tab takes an afternoon. Getting it onto the public phone network — with a carrier, a number, DLT registration, TRAI calling windows, barge-in tuned for a noisy mobile line, retries, recordings and CRM write-back — is a quarter of engineering. Vomyra ships that half already built, so the model choice is a dropdown.
What you get, stated plainly enough to check.
Four steps, none of which involve writing telephony code.
Select it as the model. Vomyra handles session management, audio framing and reconnection — none of which you write.
Wire your CRM, calendar, payment link or REST endpoint as tools. The model calls them while the caller is still on the line.
Phone audio is noisier than a headset. Vomyra's telephony-tuned interruption thresholds stop the agent cutting itself off on background noise.
Pair with an Indian mobile-format number or your own SIP trunk, and the agent is on the network.
The agent checks availability, books the slot and sends the payment link while the caller is still talking — not in a batch job afterwards.
Default interruption settings are built for clean microphones. Vomyra retunes them for mobile lines, background noise and half-duplex handsets.
Numbers, carriers, DLT, TRAI windows, retries, recordings and CRM write-back all ship with the platform. You bring the prompt.
On Vomyra, select GPT Realtime as your agent's model and attach a phone number. Vomyra owns the telephony side — carrier, number provisioning, DLT registration, TRAI calling windows, session handling and reconnection — so the deployment is a configuration rather than a build.
Yes. Vomyra runs OpenAI's realtime speech-to-speech models on Indian telephony in production, including Hindi and code-mixed Hinglish calls, on both managed Indian numbers and bring-your-own SIP trunks. The current production model is gpt-realtime, which superseded the earlier gpt-4o-realtime preview at general availability in August 2025.
Yes. The Realtime API supports function calling mid-session and, since general availability, remote MCP servers as well. Vomyra exposes your CRM, calendar, payment provider or any REST endpoint as callable tools, so the agent can check availability, write a record or send a payment link while the caller is still on the line.
Yes. Bring your own key and pay OpenAI directly for inference while Vomyra charges only the platform fee. The managed option, where model access is included on your plan, is what most customers pick.
It supports native barge-in, so a caller can cut in and the agent stops speaking and listens. Vomyra retunes the interruption thresholds for telephony audio specifically, because settings calibrated for a clean headset cause the agent to stop on background noise from a mobile line.