Integrated voice API or build your own stack?

What am I actually choosing?

Mostly, who gets woken up when a call breaks. A call needs recognition, turn detection, a model, a voice, a phone line and code to tie them together, and when one fails, someone has to find which.

We sell all three: whole voice agents, a Realtime API, and our own models as separate APIs. Voice agents cost ₹1 a minute on the volume plan.

	Hosted voice-agent API	Realtime API under your orchestrator	Stack you assemble
Components	The provider’s stack	The provider’s stack for each turn	Your choice for each part
Turn-taking and barge-in	Provider runs it	Provider runs it; your client stops playback	You build it
Telephony	Provider dials and answers, or you bring a number	You bring the transport	You bring the transport
Tools	Configured on the agent	Your code answers function calls	Your code
Observability	The provider’s events, logs and transcripts	Provider turn metrics plus your own traces	Tracing across every service you run
Upgrades	Provider ships them	Provider ships them	You test each component change
Billing	Per minute	Per minute	Each provider’s own unit, plus hosting and your time

That’s a comparison of designs, not a benchmark.

When should I buy a hosted agent API?

When your product is the work around the call: who to ring, what the agent needs to know and what happens after. You write the prompt, pass customer data as variables and act on the webhook. We run the call.

Our voice-agent API handles phone calls, browser calls and campaigns, with single-prompt or node-based agents on every plan. Check the docs for what you need: transfer to a person, for one, isn’t available yet.

When does the Realtime API make sense?

When you already run your own bot, often on Pipecat, and want to stop operating three separate speech services. The Realtime API speaks OpenAI’s Realtime protocol over one WebSocket. Your transport, tools and context aggregators stay as they are.

You give up choosing each piece. If one of ours fails mid-session, a third-party fallback covers that step at the same price, and the turn’s metadata names who served it. Filter those turns out before you compare latency.

When should I build it myself?

When a specific component, or custom conversation behaviour, is the product, and your team will own cancellation, buffering, reconnects and end-to-end tests for every combination you ship.

You can still use parts of ours. Fourier TTS works in any OpenAI-compatible client, the Pipecat package handles streaming, barge-in and phone pacing, and transcription takes recordings from any dialler.

Leave time for adapters. Two APIs with the same request shape can still differ on voice names, sample rates and formats. Pin your client versions and test the audio you actually play.

How do I compare the cost?

Price the same workload with the same things included. A model’s per-token price is one slice; add recognition, the voice, the phone line, hosting and engineering time. For a hosted API, check the phone line, how call time is rounded and what commitment applies.

We bill per minute at your plan’s rate, in 10-second blocks, and every company starts on T5 with no commitment. The pricing guide works through examples.

How do I decide without guessing?

Run one real conversation through each option, with the same language, task, audio path and concurrency, timed the same way. The latency guide lists what to agree on first. Note whether the task succeeded, how barge-in behaved and how long your team took to explain a failed call.

Then break something on purpose. Pull a dependency mid-call or replay a webhook, and check that no business action happens twice. Our post on a six-hour provider outage shows the kind of failure worth rehearsing.