Realtime Voice: the most natural AI voice, now in Call2Me

There are two ways to build a voice AI. The one almost everyone uses — including Call2Me until now — chains three specialized systems together: speech-to-text turns what the caller says into words, a language model decides what to reply, and text-to-speech turns that reply back into audio. It works, it's fast, and it's economical. We call it the cascade pipeline, and it stays the default.

But there's a newer way. OpenAI's GPT Realtime model works directly with audio — it hears the caller and speaks back without ever converting to text in between. No transcription hop, no synthesis hop. Just voice in, voice out, from a single model that was trained to converse.

Today, that's available in Call2Me as a per-agent option we call Realtime Voice.

Why speech-to-speech sounds more human

When you remove the text bottleneck, a few things get better at once:

• Latency drops. There's no waiting for one system to finish before the next starts. The model responds while it's still processing, the way a person does.
• Interruptions feel natural. Server-side turn detection lets the AI notice when you start talking and yield, instead of talking over you or freezing.
• Tone carries through. Because the model works with audio end to end, the delivery is warmer and less robotic than stitching a synthesized voice onto generated text.

The difference is most obvious on a phone call, where every extra hundred milliseconds and every awkward overlap is something the caller feels.

When to use Realtime — and when to keep the standard pipeline

Realtime isn't a wholesale replacement. Both engines live side by side, and you pick per agent:

Reach for Realtime when the conversation quality is the product — a concierge line, a high-value sales or booking agent, a demo you want to wow people with.

Stick with the standard pipeline when you want the most economical option, or you're leaning on Turkish-proven reliability and the full feature set at the lowest cost. It supports functions, knowledge base retrieval, and call transfers, and it's what most of our production traffic runs on.

What it costs

Realtime adds a premium of $0.25 per minute on top of the base voice rate. That reflects reality: the underlying model is 2-3× more expensive to run than the cascade. Because it's optional and set per agent, you only pay the premium where you actually turn it on — everything else stays at the base rate.

Turning it on

In your dashboard, open an agent and either pick a Realtime (GPT) voice from the voice picker — which switches the engine automatically — or set the engine to Realtime yourself. A +$0.25/min badge makes the cost clear before you save. Realtime also locks transcription to the agent's configured language, so a Turkish agent stays in Turkish instead of drifting to another language mid-call.

Try it on a test agent first, listen to a couple of calls, and decide where the extra naturalness earns its keep. The standard pipeline will be waiting if you'd rather keep an agent economical — the choice is yours, agent by agent.

Frequently asked

Q.What is Realtime Voice in Call2Me?

Realtime Voice is a speech-to-speech engine powered by OpenAI's GPT Realtime model. Instead of the classic pipeline that transcribes speech to text, runs a language model, then synthesizes speech, the AI hears the audio and speaks back directly. The result is a noticeably more natural, human-like voice with lower latency and smoother handling of interruptions.

Q.How is it different from the standard voice pipeline?

The standard (cascade) pipeline chains three specialized systems — speech-to-text, an LLM, and text-to-speech. It's excellent, economical, Turkish-proven, and supports every feature (functions, knowledge base, transfers). Realtime collapses those steps into a single model that works directly with audio, which sounds more natural but costs more. You choose per agent.

Q.How much does Realtime Voice cost?

Realtime adds a premium of $0.25 per minute on top of the base voice rate, because the underlying model is 2-3× more expensive to run. It's optional and set per agent, so you only pay the premium on the agents where you turn it on. The standard pipeline stays at the base rate.

Q.How do I enable Realtime Voice for an agent?

In your Call2Me dashboard, open an agent and choose a Realtime (GPT) voice from the voice picker, or switch the engine to Realtime directly. Picking a realtime voice sets the engine automatically. You'll see the +$0.25/min badge so the cost is clear before you save.

Q.Does Realtime work well in Turkish and other languages?

Realtime locks transcription to the agent's configured language, so it stays in Turkish (or whichever language you set) instead of drifting. That said, the standard pipeline remains the most battle-tested choice for Turkish. We recommend trying Realtime on a test agent for your language and use case before rolling it out widely.