Integrated voice API or build your own stack?

What are you choosing to own?

Mirai sells at all three levels: voice agents, a Realtime API, and its own speech synthesis, speech recognition and language models as separate APIs. Voice agents run from ₹1 a minute at volume (T1, with a ₹1,00,000 monthly commitment) and ₹5 a minute pay as you go (T5). The decision itself is mostly about who carries operational responsibility. A phone conversation needs speech recognition, turn detection, a language model, speech synthesis, telephony and the code that coordinates them. When a call goes wrong, someone has to find out which of those failed.

	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

These are architecture trade-offs. They are not scores from a benchmark, and any provider’s specific capabilities still need checking.

When does a hosted agent API fit?

It fits when the valuable part of your product is the business workflow around the conversation: who to call, what the agent needs to know, and what happens after. Your team writes the prompt, passes customer data as variables and acts on the webhook, and the platform runs the call.

Mirai’s voice-agent API covers phone calls, browser calls and campaigns, with single-prompt or node-based agents on every tier. Before you depend on it, check the features you need against the docs. Transfer to a human, for example, is on the roadmap and not available today.

When does the Realtime API fit?

It fits when you already run your own bot, often on Pipecat, and want to stop operating separate recognition, model and speech services. The Realtime API speaks the OpenAI Realtime protocol. You stream caller audio over one WebSocket and receive the agent’s audio back, with your own transport, tools and context aggregators left as they are.

What you give up is the choice of each component. Every session runs Mirai’s own recognition, model and speech. If one of them is unavailable mid-session, that step moves to a third-party fallback at the same price, and each turn’s metadata names the provider that served it. Read that metadata in your traces so you can separate fallback turns from normal ones when you compare latency.

When does an assembled stack fit?

Build the pipeline yourself when a specific component or custom conversation behaviour is the product. That freedom is worth it if your team is ready to own cancellation, buffering, reconnects and end-to-end testing for every combination you ship.

You can still use parts of Mirai. Mira TTS works as the voice in any OpenAI-compatible client, and the Pipecat package handles streaming, barge-in and phone audio pacing. Offline transcription works on recordings from any dialler.

Budget time for adapters. A matching request shape can hide different defaults for voice names, sample rates or output formats. Pin the client versions you deploy and test the audio you actually play.

How should you compare costs?

Compare the same workload with the same inclusions. A model’s per-token price is only part of what a call costs. Add recognition, speech, telephony, hosting and the engineering time to run them. For a hosted API, check whether telephony is included, how call duration is rounded and what commitment applies.

Mirai bills agent calls and Realtime sessions per minute at your tier’s rate, in 10-second blocks rounded up. Every company starts on T5 with no monthly commitment. The pricing guide lists the current rates and works through billing examples, and the AI calling page puts them in business terms.

How do you decide with evidence?

Run one representative conversation through each candidate, with the same language, task, audio path and concurrency. Use one timing definition for all of them; the latency guide lists the boundaries to agree on first. Record whether the task succeeds, how barge-in behaves and how long it takes your team to explain a failed call.

Add a failure drill. Take a dependency away mid-call or replay a webhook, and check that the application recovers without repeating a business action. Mirai’s journal post on a six-hour provider outage shows the kind of failure worth rehearsing. Choose the architecture that meets your product’s requirements with the team you have.