Introduction

Chariot, the Delhi-based speech lab backed by the IndiaAI Mission, released Chariot-1-TTS-Multilingual on November 12. It's a closed text-to-speech model for 11 Indic languages plus English, with 50+ voices, code-mixing between any two supported languages in the same sentence, and instant voice cloning. You can use it today in Chariot's web app and Studio, or through the API as model_type: "chariot-multilingual-v1". There are no open weights.

The pitch is reading in context. Instead of you writing normalization rules, the model decides how to say things like ₹1,250, Chapter IV, or B+, which it reads as "B positive" for a blood group and "B plus" for a grade.

Sources: Chariot's launch post (dated November 12, 2026), the Chariot TTS page, the API models guide, and the voice cloning page. Coverage: Analytics India Magazine.

What Shipped

  • Name and maker: Chariot-1-TTS-Multilingual, by Chariot Artificial Intelligence Pvt Ltd (Delhi). Chariot says its founders came from Google DeepMind and Meta FAIR.
  • Modality: text-to-speech. Supported languages are Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu, Urdu and English (American, British and Indian accents).
  • Code-mixing: any two supported languages can appear in one sentence, so Hinglish, Tanglish and similar mixes work.
  • Voices: a library of 50+ voices. Chariot says every library voice is a consented clone of a real speaker.
  • Length: up to 1,000 characters per request, for both sync and streaming.
  • Access: closed model. You can use it in the web app, Chariot Studio, or the API. It isn't on Hugging Face and isn't available for self-hosting.
  • Data residency: Chariot says TTS inputs and generated audio are processed in India.

How to Call It

  • Endpoints: POST https://api.chariot.in/v1/tts returns a WAV file. POST /v1/tts/stream streams audio over HTTP, and wss://api.chariot.in/v1/tts/ws handles WebSocket streaming.
  • Model switch: pass "model_type": "chariot-multilingual-v1" in the JSON body. If you leave it out, you get the older chariot-v0, which has a 700-character sync limit.
  • Auth: send your key in the chariotai-api-key header. On WebSocket you can use an api_key query parameter instead.
  • Voices: get a voice UUID from GET /v1/voices, and pick one whose models array includes chariot-multilingual-v1.
  • Speed: v1 supports rates other than 1.0, but only the values listed in that voice's models[].speeds. Any other rate returns HTTP 422.
  • Audio: streams are raw 16-bit PCM at 44.1 kHz mono. Chariot quotes a 75 ms median time to first audio over WebSocket, measured in its own October tests.

Voice Cloning and Plans

  • Clone input: 10 to 60 seconds of one speaker (about 30 seconds works best), as WAV, FLAC, MP3 or M4A up to 20 MB. Chariot says a clone is ready in about 20 seconds.
  • Cross-lingual: a voice cloned from Hindi speech also speaks English, Tamil and the other supported languages.
  • Consent: you have to confirm you have the rights to the voice, and Chariot's terms say you must be able to prove that consent if asked.
  • Plans: Free gets 10,000 credits and the voice library, but no cloning. Starter is $5 (₹499) a month for 100,000 credits and 3 clones. Startup is $49 for 1.5M credits and 50 clones. Scale is $299 for 10M credits and 500 clones. Commercial rights come with paid plans only.

Numbers (Self-Reported)

All of these come from Chariot's own evaluation, dated November 11, 2026. It covers 8,655 text cases and runs rival models through their public APIs. No one has reproduced the results independently yet.

  • Hindi character error rate: 7.60% across 644 prompts, averaged over three ASR judges. ElevenLabs v3 scored 8.00%, Cartesia Sonic-3.5 8.06%, Sarvam Bulbul v3 8.21%, AWS Polly 8.40% and OpenAI TTS-1 8.90%.
  • Nine-language Indic CER: 6.47% three-set mean, against 6.87% for Sarvam Bulbul v3 and 8.38% for ElevenLabs v3. Chariot also reports the lowest average in eight of the nine individual languages.
  • English verbalization accuracy: 93.78% on 1,029 sentences across 42 categories (currency, dates, URLs, chemical formulas, tickers). That narrowly beats ElevenLabs v3 at 93.51%. Cartesia Sonic-3.5 scored 89.92%.
  • Plain-English NISQA: 4.901 on 1,088 Seed-TTS-eval prompts, against 4.896 for ElevenLabs v3. NISQA is an automated quality predictor, not a human listening test.
  • The margins at the top are small, and the comparison set leaves out Google Cloud TTS.

Caveats

  • No open weights, no model card and no parameter count. It's an API product.
  • The 11-language list is from the launch post. Chariot's TTS page also advertises "15+" languages, including German, Spanish and Chinese, so check which voices support which languages in GET /v1/voices.
  • The 1,000-character cap means long-form narration has to be chunked into paragraphs. Chariot Studio handles multi-voice scripts with music for you.
  • Every benchmark is vendor-run. Test it on your own Indic copy, especially numbers and code-mixed lines, before you put it in production.

Conclusion

Chariot-1-TTS-Multilingual is a closed, API-first Indic TTS model with three selling points: free code-mixing between any two of its 12 languages, cross-lingual voice cloning from about 30 seconds of audio, and pronunciation that reads prices, dates and symbols from context. It's a direct competitor to Sarvam Bulbul and ElevenLabs v3 on Hindi. It also pairs with the Chariot-1-ASR Realtime model the lab launched on October 22. Start with the free 10,000 credits and a few of your own tricky code-mixed sentences.

—Anabel ♡