Voice AI APIs for developers building in India.

Which one do I start with?

Pick the part of the call you’d rather not build. All four APIs share one sk_live_ key and one prepaid wallet. Voice agents cost ₹1 a minute on the volume plan.

Your application needs	Start with	What your team still owns
An agent that rings customers, or answers your number	Voice-agent API (/v2/agents, /v2/calls, /v2/campaigns)	The prompt, the customer data you pass in, and what happens after the call
Your own app or Pipecat bot, with us doing recognition, the model and speech	Realtime API (OpenAI Realtime protocol over WebSocket)	Audio transport, tools and session settings
Hindi or Hinglish speech from text	Fourier TTS (/v1/audio/speech and a streaming WebSocket)	Text generation, playback and the conversation itself
Transcripts of calls that already happened	Offline transcription (/v2/transcriptions)	Storage and review of the transcripts

Streaming recognition runs inside every agent and Realtime session. On its own it’s in private preview, so ask before you design around it. The LLM is priced in the docs, and we switch it on when you ask.

Agents speak Hindi, English, Gujarati, Tamil and Telugu, and Hinglish, in four voices: ashu, neha, shruti and sameer.

What happens when I place a call?

POST /v2/calls answers at once with 202 Accepted, a call ID and status: "queued". The phone hasn’t rung yet. Everything after that arrives as signed webhooks.

1. Create a key under Developers in the sandbox. You see the secret key and a whsec_ webhook secret once, so keep both on your server.
2. Create an agent with POST /v2/agents: a name, a system prompt, an opening line, a voice and a language such as hi-IN.
3. Place the call with POST /v2/calls and an Idempotency-Key from your own record, so a retry never rings anyone twice.
4. Save the call ID before you return. A webhook can land while the request is still in flight.
5. Verify X-Mirai-Signature over the raw body, dedupe on the event id, and update your records from the final event.

The Python phone-call guide and the webhook guide have runnable code. For a whole list, use campaigns.

Calls to https://sandbox.voice.miraiminds.co bill at T5. After your company completes Go live in the console, send the same requests to https://prod.voice.miraiminds.co. Before that, prod answers 403 production_workspace_required.

What should I test with Indian callers?

The way your callers really talk. Try a Hindi sentence with an English product name, an amount and a local address. Add someone who corrects a digit halfway, and someone who talks over the agent. Listen on a real handset, because carrier audio isn’t browser audio.

Write prompts with Hindi in Devanagari, English words in Latin script and amounts as words; the prompting guide explains why. For voice quality, see the Hindi TTS guide. For timing, see the latency guide.

Where does the audio go?

Calls run on our own GPUs in India: recognition, the voice and the model. If one of our engines is slow, a backup provider in India takes that turn. Recordings are stored in Finland until our move to storage in India. The terms list exactly where call data goes.

Your phone provider, analytics and tools are yours to map. We keep recordings and transcripts until you ask us to delete them.

How do I know it’s ready for production?

Follow one real job end to end. Does the caller understand the opening line? Does the agent recover from a corrected order number? Does the final event reach the right record, and can GET /v2/calls catch you up if a webhook is late?

A completed call isn’t a completed task. status: "completed" only means the call connected and ended normally. Whether the customer said yes comes from post-call analysis or your own review. When something breaks, GET /v2/agents/{id}/logs shows tool errors, tools the model asked for that the agent doesn’t have, and the variables each call received.

Before you forecast spend, read the pricing guide. To decide how much of the stack to own, compare integrated and assembled stacks.