Voice AI infrastructure in India: TTS, STT, LLM and agent APIs

What are the layers of a voice AI stack?

A phone call with an AI agent passes through four pieces of software. Speech to text turns the caller’s audio into words. A language model decides what to say. Text to speech turns the reply back into audio. An agent platform runs the call around them: the phone line, turn-taking, interruptions, tools, and what happens after the call ends.

You can buy each layer from a different company and connect them yourself, or buy a platform that bundles them. Stitching lets you pick every component. A bundle gives you one bill and fewer network hops between the caller and the reply. Two questions narrow the choice quickly: does it speak the way your callers speak, and where does the audio go?

Who sells each layer?

This table lists what each company published on the pages we read on 16 October 2026. The comparison guides for text to speech and speech to text have the prices in full, each linked to its source.

Layer	Who sells it	How it is priced
Text to speech	Voice Infra (Mira TTS), Sarvam (Bulbul), ElevenLabs, Google Cloud, Azure, Amazon Polly, OpenAI, Cartesia	Per character for most; per token for OpenAI’s and Google’s newest models; monthly credits on Cartesia
Speech to text	Voice Infra (Mira Transcribe for recordings), Sarvam (Saaras), ElevenLabs (Scribe), Deepgram (Nova-3), Google Cloud, Azure, Amazon Transcribe, OpenAI	Per minute or per hour of audio
Language model	Voice Infra, Sarvam (Sarvam 105B and hosted open models), OpenAI, Google, Azure	Per million input and output tokens
Voice agents	Voice Infra, Sarvam Voice Agents, ElevenLabs, Deepgram Voice Agent API, Cartesia Managed Agents, OpenAI GPT-Live, and agent platforms such as Vapi, Retell AI, Bland AI and Bolna	Per minute of call, with telephony and model usage included or billed on top depending on the vendor

On agents, the published per-minute figures we read today are: Sarvam, “as low as ₹3 a minute” with the models included; ElevenLabs, $0.08 a minute for its Speech Engine; Deepgram, $0.075 a minute on the standard pay-as-you-go tier of its Voice Agent API; Cartesia, $0.06 a minute plus $0.014 a minute for a Cartesia phone number; and OpenAI, $0.05 a minute for GPT-Live sessions with backend model and tool usage charged separately. Our platform comparison lists Vapi, Retell AI, Bland AI, Bolna, Gnani.ai and Skit.ai as each published them on 8 October 2026.

Which providers publish an India data location?

Data location is where pages differ most, and where the wording matters. Storage in a region is not the same as processing there.

Provider	What its own pages say
Voice Infra	Audio, speech recognition, TTS and the model run in India (Mumbai, Noida and Delhi). The control plane (workflow orchestration and the API) runs in Europe, off the audio path
Azure Speech	Central India supports neural text to speech and real-time, fast and batch transcription
Amazon Polly	Neural voices run in Asia Pacific (Mumbai), among other regions
OpenAI	The India region lists storage, not processing, for the audio endpoints, and needs approval through sales
ElevenLabs	An India isolated environment for storage, for Enterprise customers; processing may still happen outside it
Google Cloud	TTS endpoints are global, US and EU, with Chirp 3: HD marked for Tokyo and Singapore and Mumbai listed without a model; Hindi speech to text is listed in US, EU and Singapore regions
Deepgram	A dedicated EU endpoint; the pricing page names no India region
Cartesia	The pricing page names no hosting region
Sarvam	The pricing and model pages we read do not name a hosting region

If your compliance team needs a firm answer, ask each shortlisted vendor in writing, and name the exact API you will call.

Where does Voice Infra sit?

We sell all four layers on our own models, and every layer draws on one rupee wallet with one sk_live_ key.

Mira TTS is the voice our agents use, offered as a standalone API at ₹1.80 per 1,000 characters through OpenAI’s speech protocol. Mira STT is the recognition model inside every agent, a fine-tune of the open SraVaani-1.0 model; recorded calls can be transcribed through the public API at ₹18 an hour of audio, and a standalone streaming API is in private preview. Our LLM is priced at ₹14.38 input and ₹57.52 output per 1,000,000 tokens, with 1,000 requests a minute on pay as you go, and you contact us to switch it on; a standalone endpoint is on the roadmap.

Phone agents run the whole stack: ₹1 a minute on T1 with a ₹1,00,000 monthly commitment, ₹3 on T3 with ₹50,000 a month, and ₹5 pay as you go on T5, all before GST and billed in 10-second blocks. T5 and T3 include Mirai phone calling; your own telco, WhatsApp or browser calling works on every tier.

There are gaps to plan around. Today we speak Hindi, English and Hinglish, and other Indian languages are still on the roadmap. Transfer to a person is not available yet. If you need either now, the companies in the tables above cover them.

How should you choose?

Start from the calls you will make. If they are in Hindi with English mixed in, shortlist the vendors whose voices and recognisers say they handle that, then test them on your own recordings and scripts; Voice Arena’s leaderboards and our benchmark tables are a starting point, not a verdict. If you already run an orchestrator such as Pipecat, buy the layers. If you want calls placed this week, an agent platform is less work. Check the data location for the exact API, and price the whole minute, including the phone line and the model, before you compare per-minute rates.