Chariot-1-TTS-Multilingual Explained: 11 Indian Languages, 50+ Voices, Code-Mixing and India Pricing
Illustrative image: a studio microphone. Chariot-1-TTS-Multilingual is a software model, so no product photograph exists. Image: Wikimedia Commons / Shixart1985, CC BY 2.0
Key Highlights
- Launched 12 November 2026 by Chariot, a Delhi frontier AI lab and one of the first four teams selected under the IndiaAI Mission.
- Speaks 11 Indic languages plus English: Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu and Urdu, with 50+ voices.
- Code-mixing works across any two supported languages in the same sentence, and a voice can be cloned from a few seconds of reference audio.
- Context-aware reading: ₹1,250 is spoken as "one thousand two hundred and fifty rupees", and "B+" is read differently as a blood group and as a grade.
- In Chariot's own evaluation of 11 November 2026, the model ranked first on Hindi intelligibility (7.60% character error rate), English verbalization accuracy (93.78%) and plain-English audio quality (NISQA 4.901).
- Free to start with 10,000 credits, about 10 minutes of audio, no card required; paid plans start at ₹499 a month.
What happened?
Chariot, a frontier AI research lab headquartered in Delhi, launched Chariot-1-TTS-Multilingual on 12 November 2026. The lab was founded by Parth Sarthi and Suvrat Bhooshan, former researchers at Google DeepMind and Meta's FAIR lab, and was previously known as Gan.ai. It is one of the first four teams selected under the IndiaAI Mission, the government's ₹10,372 crore programme launched in March 2024 to build India's own foundation models.
The new model is the lab's second public release, following Chariot-1-ASR Realtime, a speech-recognition model for 10 Indian languages launched on 22 October. Where ASR listens, the new model speaks: it turns written text into natural-sounding speech, and it is available now on the Chariot platform at platform.chariot.in and through the company's API.
Chariot describes its longer-term direction as "speech reasoning": models that understand not just the words in a sentence but the timing, emphasis and intent behind them. This release is presented as the first public step in that direction for speech generation.
What Chariot-1-TTS-Multilingual can do
| Capability | Detail |
|---|---|
| Languages | 11 Indic languages plus English: Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu, Urdu and English |
| Voices | 50+ voices, including native voices for each Indian language and American, British and Indian English accents |
| Code-mixing | Any two supported languages can be mixed in the same sentence, including Hinglish and Roman-letter Hindi |
| Voice cloning | Clone a voice from a few seconds of reference audio; the cloned voice speaks every supported language |
| Max length | Up to 1,000 characters per generation |
| Access | Web platform, REST API, HTTP streaming and WebSocket; about 75 ms to first audio over WebSocket |
The headline feature is not the language count alone. Most text-to-speech systems read text literally, so developers have to pre-process it: expand numbers, spell out abbreviations, add pronunciation hints. Chariot says this model makes those decisions itself, from context, so a developer can send the text as written and get speech that sounds like a person reading it.
How it reads text the way people say it
Chariot's announcement gives four illustrative examples of the judgement calls the model makes. The same written form can need a different spoken form depending on what surrounds it:
| Written expression | Appropriate spoken form |
|---|---|
| ₹1,250 | One thousand two hundred and fifty rupees |
| Chapter IV | Chapter four |
| Blood group B+ | Blood group B positive |
| A grade of B+ | A grade of B plus |
This is the practical problem the model is built for. Indian text constantly mixes scripts, languages, numbers and abbreviations: a payment reminder with an amount, a date and a UPI handle; a classroom narration that switches from Hindi to English mid-sentence; a helpline message with a ticket number and a URL. Getting those readings wrong is what makes synthetic speech sound robotic or, worse, confusing.
The benchmarks, and one important caveat
Chariot tested the model on three tiers: intelligibility (does the speech faithfully carry the intended words), reasoning accuracy (does it interpret written expressions correctly) and audio quality (how human the result sounds to an automated evaluator). The headline numbers come from Chariot's own evaluation of 11 November 2026, covering 8,655 test cases, and the company says it ranked first on all three headline metrics among the configurations it compared.
Illustrative image: a labelled audio waveform. Image: Wikimedia Commons / ItzaGarciaOrdonez, CC BY-SA 4.0
| Metric (higher or lower is better) | Chariot-1-TTS-Multilingual | Closest rival tested |
|---|---|---|
| Hindi character error rate (lower is better) | 7.60% | ElevenLabs v3, 8.00% |
| Nine-language Indic error rate (lower is better) | 6.47% | Sarvam Bulbul v3, 6.87% |
| English verbalization accuracy (higher is better) | 93.78% | ElevenLabs v3, 93.51% |
| Plain-English audio quality, NISQA out of 5 (higher is better) | 4.901 | ElevenLabs v3, 4.896 |
The caveat matters. These are Chariot's own tests, run against selected versions of rival systems, and the coverage varied by benchmark: not every rival was tested on every measure. The margins on audio quality are also thin: 4.901 against ElevenLabs v3's 4.896 is effectively a tie, while the intelligibility gap is clearer. Independent, third-party comparisons have not been published yet, so treat the "ranked first" claim as the company's own result, not a settled verdict. What the numbers do show is that an Indian lab's model is now measuring itself against the global leaders on Indian-language speech, and landing close to or ahead of them on its own test suite.
Price and how to try it
Usage is metered in credits at one credit per input character, so a 200-character request costs 200 credits and about 1,000 credits buys a minute of audio. Every account starts with 10,000 free credits, roughly 10 minutes, with every voice and language included and the first five downloads free, and no card is needed to sign up.
| Plan | Price per month | Credits (approx. audio) |
|---|---|---|
| Free | ₹0 | 10,000 (about 10 minutes) |
| Starter | ₹499 | 100,000 (about 100 minutes) |
| Creator | ₹999 | 250,000 (about 250 minutes) |
| Startup | ₹4,999 | 1,500,000 (about 25 hours) |
| Scale | ₹29,999 | 10,000,000 (about 166 hours) |
Paid plans add unlimited downloads, commercial usage rights and voice cloning: 3 cloned voices on Starter, 10 on Creator, 50 on Startup and 500 on Scale. Annual billing gives two months free. Developers can call the API with a POST request to api.chariot.in/v1/tts using the model name chariot-multilingual-v1, and Chariot says text and audio are processed in India.
Why it matters for India readers
Most voice technology is built English-first and retrofitted for Indian languages. That shows up in the products Indians actually use: banking voice agents that stumble over a rupee amount, learning apps whose Hindi narration misreads English terms, and public helplines that only work well in one language. A model that handles 11 Indian languages, switches between them mid-sentence and reads numbers and abbreviations correctly is aimed squarely at those gaps.
It also matters because of who built it. Chariot is one of the four teams the IndiaAI Mission picked to build India's foundation models, and this is the second public model from that programme's speech work in three weeks. For developers building voice agents, customer-service systems, audiobooks or multilingual content for Indian users, there is now a homegrown option priced in rupees with a free tier, rather than only global providers priced in dollars. The test for readers is simple: the free tier covers about 10 minutes, which is enough to hear whether it handles your own sentences better than what you use today.
Should you care?
If you build voice products for Indian users, run a customer helpline, or produce audiobooks, podcasts or learning content in Indian languages, yes: try the free tier with your hardest sentences, the ones with amounts, dates and mixed languages, and compare the result with your current provider. If you only need occasional English narration, the established global tools already do that well, and you should wait for independent benchmarks before switching on the strength of Chariot's own test results. And if you use voice cloning, remember Chariot requires consent and the right to use the voice you clone: cloning a real person's voice without permission is not a feature to test.
FAQs
Which languages does Chariot-1-TTS-Multilingual support?
It supports 11 Indic languages plus English: Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu and Urdu, along with English in American, British and Indian accents. Every voice can speak every supported language, and any two languages can be mixed in the same sentence.
How much does Chariot-1-TTS-Multilingual cost?
It is free to start: every account gets 10,000 credits, about 10 minutes of audio, with no card required. Usage costs one credit per input character. Paid plans start at ₹499 a month for 100,000 credits, rising to ₹29,999 a month for 10,000,000 credits, with commercial rights and voice cloning included.
How does Chariot-1-TTS-Multilingual compare with ElevenLabs?
In Chariot's own evaluation of 11 November 2026, its model ranked first on the three headline metrics: Hindi character error rate of 7.60% against ElevenLabs v3's 8.00%, English verbalization accuracy of 93.78% against 93.51%, and a NISQA audio-quality score of 4.901 against 4.896. Those are the company's own tests, and no independent comparison has been published yet.
Can Chariot-1-TTS-Multilingual clone my voice?
Yes. You can create a voice from a few seconds of reference audio, and the cloned voice can then speak in every supported language. Voice cloning is included on paid plans, and Chariot requires consent and the right to use any voice you clone.
Takeaway
Chariot-1-TTS-Multilingual is a serious, India-first entry into multilingual speech AI: 11 Indian languages, genuine code-mixing and context-aware reading, priced in rupees with a free tier. The benchmark claims are the company's own and need independent testing, but the product is live today and free to try.
Try the free tier with a sentence that mixes Hindi and English and includes an amount or a date: that is exactly the test this model was built to pass. For context on the wider AI model race, see our explainers on Runway Praxis-1's open robot model, Anthropic's Claude Fable 5.2 and Alibaba's Qwen4-Max preview.
Sources
- Chariot — Introducing Chariot-1-TTS-Multilingual (official announcement, 12 November 2026)
- Chariot — Text to Speech (product page: voices, languages and pricing)
- Chariot API Docs — Credits & pricing
- Analytics India Magazine — Chariot Launches Multilingual TTS Model with Support for 11 Indian Languages
- Business Briefing Today — Homegrown AI lab Chariot launches multilingual speech model
Comments
Post a Comment