Hindi text-to-speech APIs compared: prices and features, October 2026

What does each Hindi TTS API publish?

Every price below is the vendor’s published list price, read on 16 October 2026 from the page in the last column, in the currency the vendor uses. Enterprise and volume contracts are not shown. “Hindi, as published” is the vendor’s own description.

Provider and model	Published price	Hindi, as published	Streaming	Source
mirai minds, Mira TTS	₹1.80 per 1,000 characters (pay as you go and Business)	Hindi, English and Hinglish; Hindi voices that code-switch into English	HTTP PCM stream; WebSocket	Quickstart
Sarvam, Bulbul V4 Flash	₹2.00 per 1,000 characters pay as you go; ₹1.85 Pro; ₹1.70 Growth and Business	Hindi, English, “enhi” code-mixed voices and other Indian languages	REST, HTTP streaming, WebSocket	Pricing, models
Sarvam, Bulbul V3	₹3.00 per 1,000 characters pay as you go; ₹2.70 Growth and Business	11 languages, 10 Indian plus English	REST, HTTP streaming, WebSocket	Pricing
ElevenLabs, Eleven v4 Turbo; Flash/Turbo	$0.04 per 1K characters	Hindi listed for v4 Turbo (90+ languages) and Flash v2.5	v4 Turbo over a WebSocket	Pricing, models
ElevenLabs, Eleven v4; v3; Multilingual v2	$0.08 per 1K characters	Hindi listed for each		Pricing
Google Cloud, Chirp 3: HD	US$30 per 1M characters; first 1M a month free	30 hi-IN voices	Bidirectional streaming	Pricing, voices
Google Cloud, Neural2; WaveNet and Standard	US$16; US$4 per 1M characters	4 Neural2, 6 WaveNet and 6 Standard hi-IN voices		Pricing
Microsoft Azure, Neural; Neural HD	$15; $22 per 1M characters	hi-IN neural voices and MAI-Voice HD voices	Speech SDK and REST	Pricing, languages
Amazon Polly, Neural; Standard	$16.00; $4.00 per 1M characters	Kajal (Neural) and Aditi (Standard), bilingual Hindi and Indian English	Audio stream per request	Pricing, voices
OpenAI, tts-1; tts-1-hd	$15.00; $30.00 per 1M characters	Follows Whisper’s language list, which includes Hindi; voices “optimized for English”	Chunked transfer encoding	Pricing, guide
OpenAI, gpt-4o-mini-tts	$0.60 per 1M text input tokens; $12.00 per 1M audio output tokens	As above	As above	Pricing
Cartesia, Sonic 3.6	Credit plans from free (20K credits a month) to $299 (8M); 1 credit is 1 character	Hindi among 44 languages; expanded support for Hindi in Latin script	Bytes, SSE and WebSocket	Pricing, model

Two details change the arithmetic. Google counts every character in the input, including spaces, newlines and SSML tags. Our responses report the characters billed in an X-Chars-Billed header, and we do not charge for a request that fails or stops part-way.

What do 10 lakh characters cost on each?

Ten lakh characters is 1,000,000, a reasonable month for a small production app. At list price, before tax and before any free tier:

Provider and model	Working	Cost
Mira TTS	1,000 × ₹1.80	₹1,800
Sarvam Bulbul V4 Flash, pay as you go	1,000 × ₹2.00	₹2,000
Sarvam Bulbul V3, pay as you go	1,000 × ₹3.00	₹3,000
ElevenLabs v4 Turbo	1,000 × $0.04	$40
ElevenLabs v4	1,000 × $0.08	$80
Google Chirp 3: HD	1 × US$30 (the first 1M each month is free)	US$30
Azure Neural	1 × $15	$15
Amazon Polly Neural	1 × $16.00	$16.00
OpenAI tts-1	1 × $15.00	$15.00
Cartesia Startup plan	1.25M credits included	$49 a month

We leave currency conversion to you, because the rate on your card or contract is the one that counts. At low volume, Google’s and Amazon’s free tiers can bring the bill to zero. Our GST is extra.

How do the voices rank with listeners?

Voice Arena plays two voices reading the same line and asks listeners which sounds better, without saying which model is which, then turns the votes into Bradley-Terry Elo ratings. On its Hindi leaderboard as pulled on 16 October 2026, across all categories, Mira TTS scored 993 ± 25 from 440 votes, 11th of 19 models, and Sarvam’s Bulbul V3 scored 959 ± 9 from 3,882 votes, 14th.

By category, Mira TTS was ahead of Bulbul V3 in general conversation (1053 against 961) and in content creation and education (1014 against 901). Bulbul V3 was ahead in customer support (984 against 971) and in media and entertainment (967 against 956). Mira TTS has far fewer votes, so its interval is wider. Sarvam’s newer Bulbul V4 Flash was not on that snapshot.

Ten models ranked above Mira TTS on the same pull, among them ElevenLabs’ new Eleven v4 at the top, Google’s Gemini TTS models, Cartesia’s Sonic-3.6 and ElevenLabs’ Eleven v3. OpenAI’s gpt-4o-mini-tts and Azure’s Dragon HD Omni ranked below it, and Amazon Polly was not on the board. Ratings move as votes arrive, so read the live leaderboard before relying on any of these places.

How should you test Hindi TTS on your own scripts?

A leaderboard averages over other people’s sentences. Your product says its own, so build a small test set from it:

1. Take twenty or thirty lines your app will speak most: greetings, confirmations, a correction, an apology.
2. Put in the hard parts: customer names, rupee amounts, dates, phone numbers, an OTP, and the English words your users say inside Hindi, such as “order”, “refund” or “delivery”.
3. Write Hindi in Devanagari and keep English words in Latin script, unless your real text arrives in Roman script, in which case test it that way too.
4. Render every line with each voice you shortlisted, keeping the model and voice name with every file.
5. Play the files to people who speak Hindi every day, in random order and without names, and ask them what sounded wrong as well as which they preferred.
6. Play the best two again over a real phone call, at 8 kHz, because narrowband audio hides some errors and exposes others.

For a live agent, also time the gap from the end of the caller’s speech to the first audio of the reply, through your own stack. The Hindi TTS guide has a starter set of test sentences, and the OpenAI-compatible TTS guide shows how to request audio from Mira TTS in a few lines.

If you are pricing a whole calling agent rather than a voice, compare per-minute plans. Ours run from ₹1 a minute on T1 with a ₹1,00,000 monthly commitment to ₹5 a minute pay as you go, with recognition, the model and the voice included.