What Does an Indian-Language Voice AI Agent Really Cost per Minute in ₹ (2026)? Bundled Platforms vs DIY STT + LLM + TTS
Short answer: a one-minute Hindi or Hinglish voice-agent call built from separate parts costs about ₹2.27 in our cost model: ₹1.20 for text-to-speech, ₹0.50 for speech-to-text, ₹0.40 for the phone line and just ₹0.17 for the LLM on DeepSeek V4 Flash through Velona. Bundled platforms with published rates charge about ₹4 to ₹8 a minute at self-serve prices, often before telephony. The exception is large committed volume: Bolna's reported enterprise rates of ₹1.75 to ₹2 a minute can beat building it yourself. The LLM is under 10% of the cost of a minute, so the voice and the volume deal matter far more than the model.
Indian businesses are replacing IVR menus with voice agents that understand Hindi, Hinglish, Tamil and other languages. The pricing is confusing, because every vendor quotes a per-minute rate that covers a different set of things. This guide prices each layer of a call separately, from published rate cards, so you can compare like with like.
Key findings
- Text-to-speech is the biggest cost. At about ₹1.20 a minute, the agent's voice is more than half the cost of a DIY call.
- The LLM is the smallest cost. About ₹0.17 a minute on DeepSeek V4 Flash and ₹0.26 on GPT-5 Mini, or 7% to 11% of the total.
- DIY ranges from ₹1.63 to ₹3.41 a minute, depending on the voice model, how much the agent talks and which LLM you choose.
- Self-serve bundled rates are roughly double. CallMissed starts at ₹4 a minute plus telephony, and Bolna's reported self-serve rate is about ₹5.5.
- USD platforms cost more in India. Vapi, Retell and ElevenLabs charge their platform fees in dollars on top of the models, which works out to about ₹7 to ₹8 a minute.
- Volume changes the answer. At 50,000 minutes a month, a negotiated bundled rate can be cheaper than DIY once you count engineering time.
How we built the cost model
This is a cost model built from published prices, not a measured bill. Every rate comes from a vendor's own pricing page or documentation, or from reporting that quotes it, checked on 12 November 2026. We modelled one connected minute of an inbound Hindi or Hinglish support call, with these assumptions:
- Four conversational turns a minute, with the caller and the agent each talking for about half the time.
- The agent speaks about 600 characters a minute, roughly 30 seconds of speech. We show 450 and 750 characters as the low and high cases.
- The LLM reads about 8,000 input tokens and writes about 300 output tokens a minute. Each turn resends the system prompt and recent conversation, so input adds up quickly. Output allows for the extra tokens Hindi needs.
- The whole minute of caller audio goes through speech-to-text.
Prices exclude GST. Dollar prices are converted at ₹93.85, the rate Velona used on 12 November 2026, with no card forex markup added.
What are the four cost layers of a voice AI call?
Every voice-agent minute pays for four things: the phone line, transcription, the LLM and the voice.
| Layer | What it does | Price used | ₹ per minute |
|---|---|---|---|
| Telephony | Carries the call over the phone network | Plivo India domestic calls, about ₹0.40 a minute [1] | ₹0.40 |
| Speech-to-text (STT) | Turns the caller's speech into text, including code-switched Hinglish | Sarvam Speech to Text, ₹30 an hour, billed per second [2] | ₹0.50 |
| LLM | Understands the caller and decides what to say | DeepSeek V4 Flash on Velona, ₹19.18 per 1M input and ₹51.47 per 1M output tokens [3] | ₹0.17 |
| Text-to-speech (TTS) | Speaks the reply in an Indian voice | Sarvam Bulbul V4 Flash, ₹20 per 10,000 characters [4] | ₹1.20 |
| Total | ₹2.27 |
Your own servers, monitoring and engineering time are not included. They are real costs, and we come back to them below.
Telephony
Plivo's India rate card lists domestic calls at about ₹0.38 to ₹0.40 a minute with a 30-second pulse [1, 5]. Rates through other aggregators often run higher: one industry comparison puts them at ₹1.50 to ₹3.00 a minute, against ₹0.60 to ₹1.20 for enterprises with direct SIP deals with Airtel, Jio, Tata or Vodafone Idea [6]. Check pulse rules, because 30- or 60-second billing makes short calls cost more than the headline rate.
Speech-to-text
Sarvam charges ₹30 an hour for speech-to-text, rising to ₹45 an hour with speaker diarisation, billed per second of audio [2]. That is ₹0.50 for a full minute. You only need diarisation if you record both sides on one channel.
Text-to-speech
Sarvam's Bulbul V4 Flash costs ₹20 per 10,000 characters and returns first audio in under 250 milliseconds, according to Sarvam [4]. Analytics India Magazine reported that this is a third cheaper than the ₹30 charged for Bulbul V3 [7]. The older Bulbul v2 costs ₹15 per 10,000 characters [2]. At 600 characters a minute, those choices cost ₹0.90 to ₹1.80 a minute, which is why the voice model is the biggest decision in the stack.
The LLM
Through Velona, our 8,000 input and 300 output tokens a minute cost about ₹0.05 on GPT-5 Nano, ₹0.17 on DeepSeek V4 Flash and ₹0.26 on GPT-5 Mini [3, 8]. Even a much bigger model would rarely be the largest line on the bill.
How much does a voice AI agent cost per minute in India?
Between about ₹1.60 and ₹8 a minute, depending on how you build it and how much volume you commit to.
| Option | What the rate covers | ₹ per minute |
|---|---|---|
| DIY, cheapest case (GPT-5 Nano, Bulbul v2, 450 characters) | All four layers | About ₹1.63 |
| DIY, our base case (DeepSeek V4 Flash, Bulbul V4 Flash, 600 characters) | All four layers | About ₹2.27 |
| DIY, richer case (GPT-5 Mini, Bulbul V3, 750 characters) | All four layers | About ₹3.41 |
| Bolna, reported enterprise rate | Negotiated for large, long-term volume | ₹1.75 to ₹2 [9] |
| CallMissed, Standard plan | LLM, STT and TTS. Telephony is extra [10] | ₹4, or about ₹4.40 with telephony |
| Bolna, reported self-serve rate | Bundled self-serve pricing | About ₹5.5 [9] |
| Vapi | $0.05 platform fee, with STT, LLM, TTS and telephony passed through [11] | About ₹6.96 with our DIY components |
| Retell AI | $0.07 for voice infrastructure and standard voices, plus LLM and telephony [12] | About ₹7.14 |
| ElevenLabs Agents | $0.08 a minute, plus LLM and telephony [10] | About ₹8.08 |
Bolna also publishes a component model: a flat $0.02 a minute platform fee on top of the STT, LLM, TTS and telephony you choose [13]. With our DIY components, that comes to about ₹4.15 a minute. CallMissed lists Expressive and Best-latency plans at ₹5 and ₹6 a minute, with the ₹6 plan rising to ₹8 for calls from 30 November 2026, and a 30-second minimum per call [10].
We could not find a published per-minute price for Sarvam's own voice agents or for Ringg, so they are not in the table.
What does it cost at 1,000 and 50,000 minutes a month?
| Option | 1,000 minutes a month | 50,000 minutes a month |
|---|---|---|
| DIY base case (₹2.27) | ₹2,270, plus your hosting and engineering | ₹1,13,500, plus your hosting and engineering |
| Bolna enterprise (₹1.75 to ₹2) | Not available at this volume | ₹87,500 to ₹1,00,000 |
| CallMissed Standard plus telephony (₹4.40) | ₹4,400 | ₹2,20,000 |
| Bolna self-serve (about ₹5.5) | ₹5,500 | ₹2,75,000 |
| Vapi (about ₹6.96) | ₹6,960 | ₹3,48,000 |
| Retell AI (about ₹7.14) | ₹7,140 | ₹3,57,000 |
When does building it yourself pay off?
- At 1,000 minutes a month, buy. DIY saves about ₹2,100 a month against CallMissed, far less than the engineering time it takes to build and run a real-time voice pipeline.
- At 50,000 minutes, it depends. DIY saves about ₹1 lakh a month against self-serve bundled rates, enough to fund some engineering. But a negotiated enterprise rate of ₹1.75 to ₹2 a minute can beat DIY outright.
- DIY wins on control, not just price. You can switch LLMs in one line, keep transcripts in your own systems, and change voices or languages without waiting for a vendor.
Industry analysis makes the same point: headline per-minute rates are hard to compare, because platform fees, telephony markups and failed or short calls change the effective price [14]. Ask every vendor what their minute includes, how they round short calls, and whether you can bring your own models.
Why do Indian languages cost more, and how do I keep latency low?
The Indian-language token tax
In our own measurement, the same text needed 1.65x as many tokens in Hindi and 2.51x in Tamil as in English on the GPT-5 family tokenizer. On DeepSeek's V3 tokenizer, Hindi needed 3.12x [15]. That affects the LLM layer only, and since the LLM is a small part of the bill, it changes the total by paise rather than rupees. Three habits keep it down:
- Write the system prompt in English, and let only the caller-facing replies be in Hindi or Tamil.
- Keep replies short. Every extra word costs LLM tokens and TTS characters, and short replies sound more natural on a call.
- Cache the fixed prompt. The system prompt repeats on every turn. On Velona, cached tokens appear in the response usage and are billed at the model's cached rate [8].
Latency
People notice gaps on a phone call. The pipeline adds up: STT finishing the caller's sentence, the LLM's time to first token, and TTS producing first audio. Bulbul V4 Flash's under-250 ms first audio [4] covers the last step. For the LLM, stream the reply and send each sentence to TTS as soon as it is complete, instead of waiting for the full answer. Measure time to first token on your own traffic, because it varies by model, provider and time of day.
Python example: streaming the LLM step with FastAPI and Velona
This WebSocket endpoint receives a caller's transcript after each turn, streams the LLM reply from Velona, sends each finished sentence to your TTS, and reports time to first token and LLM cost. Plug in your own STT and TTS where the comments say.
import os
import time
from fastapi import FastAPI, WebSocket, WebSocketDisconnect
from openai import AsyncOpenAI
client = AsyncOpenAI(api_key=os.environ["VELONA_API_KEY"], base_url="https://velona.in/v1")
app = FastAPI()
MODEL = "deepseek/deepseek-v4-flash"
INPUT_RS_PER_M, OUTPUT_RS_PER_M = 19.18, 51.47 # Velona, 12 Nov 2026
# Keep instructions in English: they are cheaper in tokens than Hindi or Tamil.
SYSTEM = (
"You are a polite phone support agent for an Indian D2C brand. "
"Reply in the caller's language and script (Hindi, Hinglish or English). "
"Keep each reply under two short sentences. Never read out full card, Aadhaar or PAN numbers."
)
async def speak(text: str, ws: WebSocket) -> None:
# Send each sentence to your TTS (for example Sarvam Bulbul) as soon as it is ready.
# Here we just forward the text so you can test the loop end to end.
await ws.send_json({"type": "say", "text": text})
@app.websocket("/call")
async def call(ws: WebSocket):
await ws.accept()
history = [{"role": "system", "content": SYSTEM}]
spent_rs = 0.0
try:
while True:
# Your STT service sends one final transcript per caller turn.
msg = await ws.receive_json()
history.append({"role": "user", "content": msg["text"]})
history = history[:1] + history[-8:] # last 4 exchanges keep input tokens low
started = time.perf_counter()
first_token_ms = None
reply, buffer = "", ""
usage = None
stream = await client.chat.completions.create(
model=MODEL,
messages=history,
max_tokens=120,
stream=True,
stream_options={"include_usage": True},
)
async for chunk in stream:
if chunk.usage:
usage = chunk.usage
if not chunk.choices:
continue
delta = chunk.choices[0].delta.content or ""
if delta and first_token_ms is None:
first_token_ms = (time.perf_counter() - started) * 1000
reply += delta
buffer += delta
# Flush at sentence ends, including the Hindi danda.
if buffer.rstrip().endswith((".", "?", "!", "।")):
await speak(buffer.strip(), ws)
buffer = ""
if buffer.strip():
await speak(buffer.strip(), ws)
history.append({"role": "assistant", "content": reply})
if usage:
spent_rs += (usage.prompt_tokens * INPUT_RS_PER_M + usage.completion_tokens * OUTPUT_RS_PER_M) / 1_000_000
await ws.send_json({"type": "stats", "first_token_ms": round(first_token_ms or 0), "llm_cost_rs": round(spent_rs, 4)})
except WebSocketDisconnect:
print(f"Call ended. LLM cost: about Rs {spent_rs:.3f}")
Run it with uvicorn voice:app after pip install fastapi uvicorn openai. Trimming the history to the last four exchanges keeps input tokens flat as the call gets longer. If usage does not arrive with the stream, the cost line is skipped rather than guessed.
Frequently asked questions
How much does a voice AI agent cost per minute in India?
Built from separate parts, about ₹1.63 to ₹3.41 a minute in our model, with ₹2.27 as the base case. Bundled platforms with published rates charge about ₹4 to ₹8 a minute at self-serve prices. Bolna's reported enterprise rate for large volumes is ₹1.75 to ₹2 a minute.
What is the cheapest way to build a Hindi voice bot?
Use Indian speech models such as Sarvam for STT and TTS, an Indian telephony provider such as Plivo or Exotel, and a low-cost LLM such as GPT-5 Nano or DeepSeek V4 Flash. The voice (TTS) and the phone line cost more than the LLM, so optimise those first.
How much does Sarvam AI charge for speech-to-text and text-to-speech?
Sarvam lists speech-to-text at ₹30 an hour, or ₹45 with diarisation. Bulbul V4 Flash text-to-speech costs ₹20 per 10,000 characters, Bulbul V3 ₹30 and Bulbul v2 ₹15, per Sarvam's pricing pages as of November 2026.
Is Vapi or Retell cheaper for Indian calls?
Their platform fees are close: Vapi charges $0.05 a minute and Retell $0.07 for voice infrastructure and standard voices. Both add model and telephony costs, and both bill in dollars, which adds forex costs on an Indian card. Indian platforms that bill in rupees are often cheaper at self-serve volumes.
Does the LLM matter for voice AI cost?
Less than most people think. In our model the LLM is 7% to 11% of the cost of a minute. Model choice matters more for latency and answer quality than for price.
Limits of this guide
- A model, not a bill. We priced a typical minute from rate cards. Your real cost depends on how much callers and the agent talk, call length, pulse rounding and failed calls.
- Vendor rates change. Several prices here changed during 2026, and some are reported by third parties rather than taken from the vendor's own page. Confirm before you sign.
- Bundles differ. Vendors include different layers in their per-minute rate. We have noted what each includes where it is published.
- Engineering cost is excluded. Building and running a real-time pipeline takes engineering time, servers and monitoring, which can outweigh the savings at low volume.
- No live latency test. We did not measure end-to-end latency. Measure it on your own calls with the example above.
Try the LLM layer on Velona
Velona gives you one OpenAI-compatible API key for DeepSeek, GPT, Claude, Gemini and 300+ other models, with streaming, rupee pricing on every request and an optional budget cap on each key. Top up from ₹10 with GPay, PhonePe or Paytm, with no international card needed.
Create a free Velona account and test your voice agent's LLM step with a ₹10 top-up.
Disclosure: Velona publishes this guide and sells access to the LLMs in it. Velona does not sell telephony, speech-to-text or text-to-speech, and is not affiliated with the voice platforms compared here.
Last checked: 12 November 2026.
References
- Voice pricing, India, Plivo
- API pricing, Sarvam AI documentation
- DeepSeek V4 Flash price, Velona, refreshed 12 November 2026
- Bulbul V4 Flash text-to-speech, Sarvam AI
- Twilio alternatives in India 2026, Caller Digital, October 2026
- Plivo vs Exotel vs Ozonetel vs Knowlarity vs Twilio India 2026, Caller Digital, June 2026
- Sarvam AI launches Bulbul V4 Flash for real-time speech in 11 Indian languages, Analytics India Magazine
- Velona API documentation, Velona
- The unit economics of India's voice AI boom, Inc42
- Voice AI pricing, CallMissed
- Vapi vs Retell AI vs custom, Fora Soft, July 2026
- AI voice agent pricing: a practical guide to total cost, Dasha
- How much does Bolna Voice AI cost?, Bolna documentation
- Voice AI pricing in India, Caller Digital
- Does AI cost more in Hindi? We measured the token tax for 9 Indian languages, Velona, October 2026
- Cheapest models, Velona, 12 November 2026
Related: India's own AI models in 2026, anonymising data before it reaches an LLM and AI price per message in ₹.