Chariot-1-TTS-Multilingual explained: Indian-language TTS

By LLMTest Lab · Nov 12, 2026 · 4 min read voicetext-to-speechindia
On this page

On this page

  1. What it is
  2. Access and price
  3. The claims, and who made them
  4. Who should care
  5. What we will test

Not tested by us yet. Our test harness calls chat models over an API and cannot score audio, so we have not listened to this model ourselves. Everything below comes from Chariot's text-to-speech page and launch coverage in Analytics India Magazine and CIOL. We'll update this post when we can run it on a fixed set of scripts.

Chariot-1-TTS-Multilingual, launched on 12 November 2026, is a text-to-speech model built for Indian languages. It covers 11 Indian languages plus English, ships 50+ voices, mixes two languages inside one sentence and clones a voice from a short clip. If you build voice agents or helplines for Indian users, it is worth a listen. If you need an independent ranking against ElevenLabs or Cartesia, none exists yet.

Chariot-1-TTS-Multilingual
Maker Chariot (formerly Gan.ai), Delhi
Launched 2026-11-12
Languages 11 Indian languages plus English (Hindi, Marathi, Bengali, Tamil, Telugu, Kannada, Malayalam, Gujarati, Punjabi, Odia)
Voices 50+, voice cloning from a few seconds of audio
Input limit 1,000 characters per request
API model type chariot-multilingual-v1 (per the product page)
Time to first audio about 75 ms median over WebSocket (product page, Chariot's figure)
Weights Not released; hosted API only

What it is

Chariot says the model reasons about context before it speaks. Instead of rule-based text normalization, it decides how to read numbers, symbols and abbreviations from the surrounding sentence. The launch example is "B+", read as a blood group in one sentence and as a grade in another. The product page lists currencies, dates, doses and abbreviations as cases it handles without markup.

The TTS model sits on the same Chariot-1 foundation as Chariot-1-ASR Realtime, the speech recognition model the company launched on 22 October 2026. The product page also lists Hinglish, Tanglish and Hindi typed in Roman letters.

Access and price

The API has three endpoints: REST returning a WAV file, a streaming endpoint returning 44.1 kHz PCM, and a WebSocket for sentence-by-sentence voice agents. Prices on the product page are in rupees, excluding tax:

Plan Per month Audio per month Concurrent requests
Free ₹0 about 10 min 2
Starter ₹499 about 100 min 3
Creator ₹999 about 250 min 5
Startup ₹4,999 about 1,500 min 10
Scale ₹29,999 about 10,000 min 25

The Free plan carries no commercial rights. Per-character API pricing is not published, and neither is anything about licensing for the model itself. Chariot says inputs and generated audio are processed in India under the Digital Personal Data Protection Act.

The claims, and who made them

Every quality number is Chariot's own. At launch it reported a 7.60% character error rate on Hindi, 6.47% averaged across nine other Indic languages, and 93.78% accuracy on an English verbalization set of more than 1,000 sentences in 40+ categories. It says it ranked first on several measures against ElevenLabs, Cartesia, Microsoft, OpenAI, AWS and Sarvam.

The product page carries an October evaluation with named rivals. On hard written English it shows 94.0% for Chariot against 93.5% for ElevenLabs v3 and 89.9% for Cartesia Sonic-3.5. On Hindi character error rate it shows 7.73% for Chariot against 8.21% for ElevenLabs v3 and 8.30% for Cartesia. The Hindi figure differs from the launch figure (7.73% versus 7.60%), and the page does not say why.

Three caveats:

  • CIOL notes that the models compared varied from test to test, and that independent testing is still needed.
  • Chariot's page says Google Cloud TTS was not part of the comparison.
  • Latency, voice-cloning consent checks and per-character cost were not disclosed at launch.

The page also describes a wider list than the launch coverage: German, Chinese and Urdu voices, plus Assamese, Nepali and Sindhi. We treat the 11-plus-English list as the launch scope and the rest as unverified.

Who should care

  • Teams building voice agents for Indian customers. Code-mixing and in-context number reading are what Indian call flows need most.
  • Builders comparing open and hosted voices. This one is hosted only. ZONOS3-Mini is the open-weights alternative, and Underdog Woof sits at the tiny on-device end.
  • Anyone choosing a speech vendor on a benchmark. Wait for a third-party run first.

What we will test

We would read 20 scripts in Hindi, Tamil and Hinglish, with prices, dates and phone numbers, through Chariot and two hosted rivals. We would score word error rate with a speech recognizer, clone three voices and check speaker similarity, then compare time to first audio and cost per hour of audio. Until then, treat every quality figure above as a vendor claim.

Related articles

ZONOS3-Mini explained: 100M-active open TTS, 43 languages
Zyphra's ZONOS3-Mini is an Apache 2.0 text-to-speech model with 800M total and 100M active parameters, voice cloning and a browser build. Specs and claims.
Underdog Woof first look: a 32MB open TTS model on CPU
Underdog Woof is a 38M-parameter English text-to-speech model that runs on CPU in a 31.7 MB download. Specs, vendor-reported WER, license limits, who it suits.