Eight Indian languages are live on every call, and HinglishEight Indian languages live Mira STT: a final transcript 82 ms (p80) after the caller stopsTranscript in 82 ms (p80)

AI agents that call your customers, in 8 Indian languages.

Live today. ₹200 of calls free to start.

to start, phone line included: ₹5/min
all-in at volume: 1 lakh minutes, ₹1,50,000¹: ₹1.50/min
word errors on Hindi phone calls (Sarvam 21.6%)²: 18.3%

Start free

Hindi नमस्ते, मैं नेहा बोल रही हूँ। आपका पार्सल कल शाम तक पहुँच जाएगा। अगर उस समय आप घर पर नहीं होंगे, तो बता दीजिए, हम दिन बदल देंगे।

English Hi, I am Shruti from Acme Dental Clinic. Your appointment is tomorrow at eleven thirty in the morning. Shall I confirm it, or would you like a different time?

Gujarati નમસ્તે, હું આશુ બોલું છું. તમારી લોનનો હપ્તો આવતા અઠવાડિયે ભરવાનો છે. તમે કહો તો હું અત્યારે જ ભરવાની રીત સમજાવી દઉં.

Sameer Tamil வணக்கம், நான் சமீர் பேசுகிறேன். நேற்று எங்கள் புதிய வீடுகளைப் பற்றி கேட்டிருந்தீர்கள். இப்போது கொஞ்சம் பேச முடியுமா?

Telugu నమస్కారం, నేను నేహా మాట్లాడుతున్నాను. మీ పార్సెల్ రేపు సాయంత్రంలోగా మీకు అందుతుంది. ఆ సమయంలో మీరు ఇంట్లో లేకపోతే చెప్పండి, వేరే రోజుకి మారుస్తాము.

Five real Mira lines, one request each, unedited.³ The board is drawn to show scale, not a live count.

What you get

Hand over the calls your team makes every day.

Each card plays a sample call's first line, what that call ends with, and what it costs to start and at volume.

E-commerce

Agent

नमस्ते, मैं Acme Store की AI assistant Neha बोल रही हूँ। क्या मेरी बात Rahul जी से हो रही है?

Confirm every COD order before it ships

Address corrected

₹6.67 a call to start , ₹2 all-in at volume
E-commerce

Agent

नमस्ते Pooja जी, मैं Acme Fashion की AI assistant बोल रही हूँ। आपने कल रात हमारी website पर दो kurtis cart में डाली थीं, उसी के बारे में call किया।

Win back the carts that stopped at a question

Wants the checkout link

₹9.17 a call to start , ₹2.75 all-in at volume
Logistics

Agent

नमस्ते Sanjay जी, मैं Acme Store की तरफ़ से AI assistant बोल रही हूँ। आपका parcel कल deliver नहीं हो पाया, courier ने लिखा है कि आप available नहीं थे।

Save failed deliveries before they become returns

Reattempt tomorrow

₹8.34 a call to start , ₹2.50 all-in at volume
Real estate

Agent

नमस्ते Kunal जी, मैं Acme Homes की AI assistant Shruti हूँ। आपने अभी हमारी website पर Wakad project के बारे में enquiry की थी। दो मिनट बात कर सकते हैं?

Call every new lead while it's still warm

Hot lead, site visit booked

₹15 a call to start , ₹4.50 all-in at volume

It hears a Hindi phone call better, and finishes the transcript sooner.

Voice Infra and Sarvam: speech recognition on Hindi phone calls and on read speech. Lower is better on every row.
Measure	Voice Infra	Sarvam
Errors on Hindi phone calls, word error rateLower is better	18.3% 499 public phone recordings	21.6% Saaras v4
Final transcript after the caller stops, 80th percentileLower is better	82 ms 158 Hindi phone clips, from India	290 ms Realtime API
Errors on clean read Hindi, word error rateLower is better	10.1% Kathbath read speech	5.8% Saaras v4

Sarvam is better on clean read speech. The test, run on 20 October 2026 through each company's own API, is on our Sarvam comparison.

Start free · ₹200 credit Every use case

₹200 of credit, about 40 minutes of calls. Just your email.

Why Voice Infra

The problem

Most voice agents are built from six different vendors.

The phone line, noise removal, speech recognition, the language model and the voice each come from a different company. Every hand-off adds delay and cost.

The solution

We run every part of the call ourselves.

The phone line, speech recognition and language model run on our own GPUs in India, with the voice on top, for one price a minute.

The line. Theirs: Twilio / Plivo / Exotel, vendor edge, often outside India — network 1 · a separate telephony integration. Ours: Indian carrier trunks, our own interconnect, no CPaaS in between.
Orchestration. Theirs: Pipecat Cloud / LiveKit / Vapi, US region — network 2 · a TLS handshake and a vendor queue. Ours: our worker, a LAN hop, one data centre.
Noise. Theirs: Krisp, vendor cloud — network 3 · the audio leaves again. Ours: Mira denoise, in process, no network at all.
ASR. Theirs: Deepgram / AssemblyAI / Sarvam, US or Indian region, per vendor — network 4 · a separate speech recognition stream. Ours: Mira ASR, the same card as the model and the voice.
LLM. Theirs: OpenAI / Anthropic / Gemini, US region — network 5 · another connection for every reply. Ours: open source LLM, the same card.
TTS. Theirs: ElevenLabs / Cartesia, US region — network 6 · a separate speech synthesis stream. Ours: Mira TTS, the same card, first audio back on the same LAN.

Response time

Stitched stackVoice Infra

1.6s 850ms

ASR + VAD + buffer900 ms590 ms
LLM400 ms150 ms
TTS300 ms110 ms

You finish speaking → first reply audio. Stitched stage split is estimated.

Now, let’s talk about the most important aspect: price.

India needs pricing made for India.

One rupee used to buy you
a call from an STD box.

₹1.50 a minute, all-in. A voice agent.₹1 for the agent, plus our phone line and call insights. Now the machine on the line talks back.

Noise reduction

Orchestration

ASR

LLM

TTS

Mira
Stack

A silver Indian one-rupee coin dated 2026

Five products to connect.
One conversation to get right.

Every layer. One Mira Stack.

₹1.50^* all-in / min at volume

नमस्ते, मैं आशु बोल रहा हूं। आपकी EMI पांच तारीख को due है — क्या online pay करेंगे?

Rendered by Mira TTS · 2026-09-03

Start on T5 at ₹5/min with no commitment, line and insights included. *T1 at volume: ₹1/min for the agent, ₹1.50 with Mirai telco (+₹0.30) and insights (+₹0.20), on a ₹1,00,000/month standard minimum. No Mirai telco charge for BYO, WhatsApp or browser calls. All prices exclude GST.

Running in production

How we got Voice Infra into production.

Build the fleet

We started by renting GPU capacity and building our fleet in Indian data centres. Then we took charge of how that compute runs, scales and serves every call.

Rented compute. Operated by us.

Own the core

Speech recognition and the language model run on our own GPUs in India. On Hindi phone recordings, Mira STT makes fewer word errors than Sarvam’s and finishes the transcript sooner.

Hindi phone calls: 18.3% word errors, against Sarvam’s 21.6%. On clean read speech Sarvam does better.

Prove it live

Before opening to the public, we spent three months running live with beta customers. Real calls, real feedback. Test, fix, test again.

Three months with beta customers before public access.

Stay bootstrapped

Bootstrapping gives us room to focus on the core: better models, faster inference and lower costs. We build for customers, on our own terms.

Customer revenue goes back into the core.

Voice Infra, made in India

Build it with us.

Bring the voice experience you want to build. Let’s talk about how to make it happen.

Book a call

15 minutes. You, us, and your next idea.

Built for builders

All this infrastructure. One API call.

The GPU fleet, the speech models, the inference engine and all the orchestration. We bring it together and keep it running. You build the conversation.

Start a conversation POST /v2/calls

from mirai import Mirai

# Uses MIRAI_API_KEY
mirai = Mirai()
call = mirai.calls.create(
    agent_id="YOUR_AGENT_ID",
    to="+91XXXXXXXXXX",
)

curl https://api.voice.miraiminds.co/v2/calls \
  -H "Authorization: Bearer $MIRAI_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: YOUR_UNIQUE_CALL_ID" \
  -d '{
    "agent_id": "YOUR_AGENT_ID",
    "to": "+91XXXXXXXXXX"
  }'

Use an agent you’ve configured in your workspace, and the number you want to call.

Read the docs TTS quickstart

FAQs

Before you build.

How does pricing work?

Every company starts on T5: ₹5 a minute, pay as you go, with our phone line and call insights included. At volume, T1 is ₹1.50 a minute all-in: ₹1 for the agent, ₹0.30 for our line and ₹0.20 for insights, on a ₹1,00,000 monthly minimum. T3 sits between, at ₹3 on ₹50,000 a month. All prices exclude GST.

How fast does it reply?

About 850 ms on average, from the moment you finish speaking to the first reply audio: 590 ms for ASR, VAD and buffering, 150 ms for the LLM and 110 ms for TTS. It varies with speech, network and load, and the network distance to the caller is extra.

Which languages?

AI agents call in Hindi, Gujarati, Marathi, Bengali, Tamil, Telugu, Kannada and Malayalam, in speech recognition and in the voice, and in the Hinglish mix too. English works on every call.

Where does the audio go?

Speech recognition and the language model run in Indian data centres. You can review call recordings, transcripts and costs in the console. Website analytics and booking providers are covered separately in our privacy notice.

Do I need my own telephony?

T5 and T3 include Mirai phone calling and insights. T1 adds ₹0.30/minute only for Mirai-managed phone calls. BYO telco, WhatsApp and browser calls have no Mirai telco charge on any tier. Provider charges you pay directly are separate.

Can you support a larger deployment?

Yes. Talk to us about call volume, concurrency, workflows and the languages you need. We’ll work through capacity and commercial terms with you before you commit.

How do I try it?

Press Start free. The sandbox opens without a sign-up, and we ask for your email when you test a call. Once you confirm it, the workspace has ₹200 of credit, about 40 minutes of calls. You can also talk to Mira in your browser from any page.

Is there an API?

Yes — a v2 REST API with one auth header: create an agent, place a call, read the transcript, recording and cost. Mira TTS also speaks OpenAI’s /v1/audio/speech. Everything is documented at docs.miraiminds.co.

Still have a question? Talk to us. We’re the people building it. sneh@miraiminds.co