Chariot-1-TTS-Multilingual Review: 11 Indian Languages, 94% Accuracy, and ₹4.99 a Minute

A Delhi lab founded by former Google DeepMind and Meta FAIR researchers launched a text-to-speech model today that is built for the way Indians actually write. Chariot-1-TTS-Multilingual speaks 11 Indian languages plus English, mixes two languages inside one sentence, and reads a rupee amount, a date or a blood group the way a person would say it, without the text-normalisation rules other systems make you write first. It is the second model in four weeks from Chariot, after its real-time transcription system, and it arrives with a published price list in rupees and a benchmark table that puts it ahead of ElevenLabs v3 on the tests Chariot chose.

That last clause matters. Every number in the benchmark table comes from Chariot's own October 2026 evaluation, run through the public APIs of nine commercial systems. The model itself is new, the claims are company-reported, and the rival it beats most loudly, ElevenLabs, has since shipped a newer model that was not in the test. This review lays out what the model does, what each plan costs in rupees, dollars, pounds and Canadian dollars, and where the fine print sits.

What Chariot-1-TTS-Multilingual Is

Chariot-1-TTS-Multilingual is a text-to-speech model, sold as chariot-multilingual-v1 on Chariot's API. The headline feature is not the voice quality alone. It is the model's grasp of written text as people actually type it. Most text-to-speech systems need clean, expanded input: numbers written out, abbreviations spelled, dates reformatted. Chariot says its model uses context to decide how a written expression should be spoken, so the reasoning sits inside the model instead of in a pile of rules you maintain.

The examples on the product page show what that means in practice. ₹2,450 is read as a rupee amount, 15 March as a date, 06:45 as a time, 500 mg as a dose, Dr. as Doctor, and B+ as B positive, while order and flight numbers are read digit by digit. The distinction Chariot leans on is B+ itself: the same two characters mean a blood group in one sentence and a grade in another, and the model is supposed to tell them apart from context. Analytics India Magazine, which covered the launch, reports the same example and adds that the model distinguishes currency amounts, dates, abbreviations and technical expressions by context rather than by rules.

The language list is the other half of the story. The model speaks 15 or more languages: English, German and Chinese, plus Hindi, Urdu, Bengali, Marathi, Telugu, Tamil, Gujarati, Kannada, Malayalam, Odia and Punjabi. Every voice speaks every language, so one brand voice can carry a message from one market to the next. The catalogue holds more than 50 voices, including 12 native Hindi voices, 8 British English and 6 American English, 5 Indian-accent voices for Assamese, Nepali and Sindhi, and 2 native voices for each of the other Indian languages.

Code-mixing is where the model is aimed most directly at Indian users. Write Hindi in Devanagari or in Roman letters, mix in English words either way, and the voice keeps one accent across both. The same goes for Tanglish, Banglish and the other mixes of an Indian language with English. That matters because the text people actually send, in chats, support tickets and social posts, is rarely in one clean language. Voice translation is a different job from text to speech, but the two meet in multilingual products; our DeepL Aure review covers a tool that translates spoken voice while keeping the speaker's tone.

The Benchmarks, and the Caveats That Come With Them

Chariot's evaluation, run in October 2026, put its model and eight other commercial systems through the same sentences via their public APIs, judged by the same three speech recognisers and one language-model judge. The headline numbers, taken from Chariot's own product page:

The English test covered 1,029 sentences written the way text actually arrives, across 40 categories including customer service, where Chariot scored 96.0%, and technical and scientific text, where it scored 95.6%. The Hindi test used 376 native sentences plus 268 general and code-mixed sentences. Naturalness was rated clip by clip with NISQA: Chariot scored 4.91 in English, tied first with Grok, and 4.97 in Hindi, first outright. For latency, the median time to first audio over the WebSocket endpoint is 75 milliseconds, so a voice agent can start speaking before the full reply is written.

Three caveats belong next to those numbers. First, they are Chariot's own results, and no independent lab has repeated them. Moneycontrol, covering Chariot's earlier transcription launch, noted the same point about that model's claims: they could not be independently verified at the time. Second, the rival in the table is ElevenLabs v3. ElevenLabs shipped Eleven v4 in October 2026, after the evaluation ran, so the table does not say Chariot beats the current ElevenLabs flagship. Third, Google Cloud text to speech was not part of the test at all, so the biggest cloud vendor in the comparison is simply absent.

There is also a small mismatch in the launch numbers themselves. The product page reports 94.0% on the English test and a 7.73% Hindi error rate. Analytics India Magazine, reporting the launch, cites 93.78% accuracy on the same English benchmark and a 7.60% Hindi error rate, plus 6.47% across nine other Indic languages. The two sets of figures likely come from different cuts of the test data, and Chariot has not published the full methodology. Treat the product-page table as the reference version, and the announcement figures as the same story told with a slightly different test set.

What It Costs: Plans in Rupees, Dollars, Pounds and Canadian Dollars

Chariot prices in rupees, with no forex markup, and every plan bills in credits: 1 credit per input character for text to speech, so a 200-character request costs 200 credits, according to Chariot's own credits documentation. The page estimates 1,000 credits at about 1 minute of audio, and says actual duration varies with language and speaking pace. Prices exclude applicable taxes. Conversions below use November 12, 2026 market rates of 1 USD = ₹93.99, 1 USD = £0.7541 and 1 USD = CA$1.426, and are approximations.

The free tier is usable for a trial: 10,000 credits, about 10 minutes of audio, all 50-plus voices and every language, with no card required. The limits are the point, though. You can download only your first 5 generated clips, commercial usage rights are not included, and voice cloning is a paid-plan feature. The free plan is for trying the voices, in Chariot's own words, not for publishing.

Starter is where the product becomes a tool. At ₹499 a month it adds unlimited downloads, commercial rights, 3 concurrent requests and 3 cloned voices. Creator doubles the per-minute value to about ₹4.00 and raises the limits to 5 concurrent requests and 10 cloned voices. The per-minute cost keeps falling as you climb, to about ₹3.33 on Startup and ₹3.00 on Scale, but the jump between tiers is steep: Scale costs 30 times Starter for 100 times the minutes.

Two pieces of fine print deserve attention. First, the minute figures are estimates built on the 1,000-credits-per-minute rule. A language that speaks slowly, or text full of long words, will burn credits faster than the table suggests, so a heavy user should measure their own burn rate in the first month rather than trust the estimate. Second, Chariot publishes no per-character dollar price for API use outside the plans. You buy a plan and spend its credits; there is no pay-as-you-go meter to fall back on if you run out mid-month, and the page does not say whether unused credits roll over.

How the Price Compares With ElevenLabs

ElevenLabs' own pricing page, read live on November 12, 2026, lists Free at $0 with 10,000 credits a month, Starter at $6 a month with 30,000 credits (about 30 minutes), Creator at $22 a month with 121,000 credits (about 121 minutes) and Pro at $99 a month with 600,000 credits (about 600 minutes). First-month promos cut Starter to $1 until November 16 and Creator to $11, and annual billing works out to $5, $18.33 and $82.50 a month respectively, since a year costs 10 months. Prices exclude taxes. On the same 1,000-credits-per-minute rule, ElevenLabs Starter costs about $0.20 a minute, Creator about $0.18 and Pro about $0.17.

On that arithmetic, Chariot's Starter at about $0.05 a minute costs roughly a quarter of ElevenLabs Starter per minute, and its Creator at about $0.04 a minute costs roughly a quarter of ElevenLabs Creator. The comparison is not like for like in every respect. ElevenLabs supports more languages overall, its free tier and paid tiers include dubbing, sound effects and music tools that Chariot does not sell, and its Professional Voice Cloning sits on the Creator tier while Chariot includes cloning from Starter. Chariot's edge is narrower and sharper: Indian languages, code-mixing, and a per-minute price that undercuts the market leader by a wide margin. The per-minute gap is the story; the feature gap runs the other way.

One more comparison is worth making for Indian buyers. Sarvam's Bulbul v3, the domestic rival in Chariot's own benchmark table, scored 88.2% on the hard English test against Chariot's 94.0%, with a Hindi error rate of 8.95% against 7.73%. Sarvam's published per-minute pricing was not part of the material reviewed for this article, so no price comparison is made here. On accuracy alone, Chariot's table puts it ahead of the other Indian system it tested.

Using It: API, Voice Cloning and Chariot Studio

The API is a single endpoint with one key. Send a POST request to https://api.chariot.in/v1/tts with your key in the chariotai-api-key header and a JSON body naming the voice, the text and model_type set to chariot-multilingual-v1, and the response is a WAV file. Streaming returns 16-bit PCM at 44.1 kHz from about 75 milliseconds, and a WebSocket endpoint speaks sentence by sentence for live agents. Each request takes up to 1,000 characters, so a longer script is sent a paragraph at a time, with the voice and settings staying the same across requests. Speed is adjustable from 0.5x to 1.5x, and the web app exports MP3 or WAV. For a voice agent that listens as well as talks, our Mistral Voxtral 3 review covers a phone-ready model at $0.04 a minute.

Voice cloning takes 10 to 60 seconds of one speaker, recorded in the browser or uploaded as WAV, FLAC, MP3 or M4A up to 20 MB. About 30 seconds of clean speech without music or background noise gives the closest match, and the voice is usually ready in 10 to 20 seconds. A cloned voice speaks every language the model speaks, whatever language the recording was in, so a Hindi recording can narrate the English and Marathi editions of the same book. Starter includes 3 cloned voices, Creator 10, Startup 50 and Scale 500, and moving to a smaller plan never deletes a voice: voices over the new allowance pause until you upgrade.

Cloning carries the one rule that matters most. You may clone only a voice you have the right to use: your own, or someone who has given you permission. You confirm this when you upload, Chariot keeps a record of it, and its Acceptable Use Policy applies. That is a lighter gate than a full consent workflow, and it puts the responsibility on the uploader. It is also the same approach the industry has settled on, and it is worth knowing before you clone a voice that is not yours.

For longer work there is Chariot Studio, which takes a script or an idea, casts a voice for every speaker, scores the piece with music and sound that start on the words they belong with, and lets you change one line at a time. It uses the same voices and credits as text to speech. The use cases Chariot lists are voice agents and IVR, ads and promos, YouTube narration, audiobooks, e-learning and reading aloud for accessibility. A bank can use it for payment reminders, a learning platform for narration, and a public helpline for spoken information across languages, according to the launch coverage. Text-based agents are the other half of this market; our roundup of 7 AI agents you can text covers the ones that live in your messages.

Who Chariot Is, and Why This Launch Exists

Chariot is a Delhi-headquartered frontier AI research lab, previously known as Gan.ai, founded by Parth Sarthi and Suvrat Bhooshan, both with experience at Google DeepMind and Meta's Fundamental AI Research lab. It is one of four teams selected under the IndiaAI Mission to build the country's foundation models, and one of 12 organisations and consortia chosen under the mission's Foundation Models pillar, alongside Sarvam AI, Soket AI and Gnani AI. The IndiaAI Mission launched in March 2024 with an allocation of ₹10,372 crore for indigenous AI infrastructure, datasets, applications and foundation models.

The company's earlier work explains why a multilingual speech model is its second launch in a month. On October 22 it launched Chariot-1-ASR Realtime, a live transcription system for 10 Indian languages, which it said outperformed every streaming and batch system it tested. Its first public model, Chariot-1 TTS, went live in beta on July 29, 2026. The new multilingual model is the production form of that line of work, with the language coverage widened and the pricing published.

Chariot's showcase work is also worth a line, because it shows who already trusts the voices. Its site lists projects with Zomato, Doordarshan, Coca-Cola, Cadbury and HDFC Bank, including the Prime Minister's Mann Ki Baat address delivered in his voice across languages for Doordarshan, and 600,000 personalised birthday songs, each sung to a different name, for Cadbury. Those are company-showcased projects, not independent reviews, but they show the model family has already been used at national scale.

The Catches: What to Check Before You Pay

The first catch is the benchmark itself. It is Chariot's own test, on Chariot's chosen sentences, judged by recognisers Chariot picked. The margins over ElevenLabs v3 are small on English accuracy, 94.0% against 93.5%, and the naturalness gap is 0.04 on a 5-point scale. Those are real leads on the reported numbers, but they are not the kind of gap that survives a change of test set. The Hindi error-rate lead is wider, 7.73% against 8.21%, and that is the number that matters most for the model's home market.

The second catch is the free tier's export limit. Five downloads is enough to test the voices and not much more. Anyone planning to publish should budget for Starter from the start, because the free tier's lack of commercial rights means work made on it cannot be published at all.

The third is the credit meter. Billing is per input character, and the minute estimates assume 1,000 credits per minute. Long words, slow languages and repeated retakes all cost more than the table implies, and Chariot publishes no rollover rule for unused credits. A voice agent running all day will burn through Starter's 100,000 credits in about 100 minutes of speech, which is not a lot of talk time for a support line. Startup's 1,500 minutes is the first tier that fits a real deployment.

The fourth is scope. Fifteen languages is a lot for an Indian product and a small number next to ElevenLabs, which Chariot itself concedes supports more languages overall. If your audience is mostly outside India, the language list is a reason to look elsewhere. If your audience is in India, it is the reason to look here first.

Finally, the data question. Text and audio are processed in India under the Digital Personal Data Protection Act, which is a genuine differentiator for Indian businesses with data-residency requirements. You can ask Chariot not to use your content to improve its models by writing to [email protected]. That opt-out is by email, not a toggle in the dashboard, and the privacy policy notes that website and product analytics may be processed by providers outside that guarantee.

Frequently Asked Questions

What is Chariot-1-TTS-Multilingual?

It is a text-to-speech model from the Delhi lab Chariot, launched November 12, 2026. It speaks 11 Indian languages plus English, German and Chinese, offers more than 50 voices, mixes languages inside one sentence, and reads numbers, dates and prices the way a person says them.

How much does Chariot-1-TTS-Multilingual cost?

The free tier gives 10,000 credits, about 10 minutes of audio, with no card required but only 5 downloads and no commercial rights. Paid plans start at ₹499 a month for about 100 minutes, then ₹999 for about 250 minutes, ₹4,999 for about 1,500 minutes and ₹29,999 for about 10,000 minutes. Prices exclude taxes.

Is Chariot-1-TTS-Multilingual really more accurate than ElevenLabs?

In Chariot's own October 2026 evaluation it was: 94.0% of hard English sentences read correctly against 93.5% for ElevenLabs v3, and a 7.73% Hindi error rate against 8.21%. The test was run by Chariot, it compared v3 rather than the newer Eleven v4, and Google Cloud text to speech was not tested.

Can it read Hinglish and Hindi typed in Roman letters?

Yes. Write Hindi in Devanagari or in Roman letters, mix in English words either way, and the voice keeps one accent across both. The same applies to Tanglish, Banglish and other mixes of an Indian language with English.

Can I clone my own voice?

Yes, on every paid plan. Record or upload 10 to 60 seconds of one speaker, and the voice is usually ready in 10 to 20 seconds. Starter includes 3 cloned voices, Creator 10, Startup 50 and Scale 500. You may clone only a voice you have the right to use.

Where is my data processed?

Text and audio are processed in India under the Digital Personal Data Protection Act. You can ask Chariot not to use your content to improve its models by writing to [email protected].

Chariot-1-TTS-Multilingual is a narrow bet made well: a model that understands how Indians write, priced in rupees at a fraction of the market leader's per-minute cost, with its benchmarks openly attributed to its own testing. The accuracy lead over ElevenLabs v3 is small and self-reported, the free tier is a trial rather than a tool, and the credit meter needs watching. For anyone building voice products for Indian users, though, the combination of 11 Indian languages, code-mixing and ₹4.99 a minute is the strongest offer in the category today. Start on the free tier, measure your own burn rate, and let the first real script decide whether Starter or Creator fits.