Editorial illustration for Chariot Launches Multilingual TTS Model Supporting 11 Indian Languages
Chariot Launches TTS Model for 11 Indian Languages
Chariot, an AI research lab based in India, has launched a text-to-speech model that handles 11 Indian languages plus English, including mid-sentence code-mixing between any two of them. The model, called Chariot-1-TTS-Multilingual, supports more than 50 voices and can clone a new one from just a few seconds of reference audio. It generates speech from text blocks up to 1,000 characters long in a single request.
The founders, who previously worked at Google DeepMind and Meta's FAIR lab, built the system around a different approach to a stubborn TTS problem: how a model decides to pronounce ambiguous text. Instead of using fixed normalisation rules, Chariot's model reasons from context, so it can tell that "B+" means a blood group in one sentence and a grade in another. The same logic extends to currency figures, dates, abbreviations, and technical terms that read one way and often sound completely different out loud.
Chariot reported benchmark comparisons against models from ElevenLabs, Cartesia, Microsoft, OpenAI, AWS, and Sarvam, claiming the top spot on several measures, including character error rates in Hindi and nine other Indic languages. Co-founder Parth Sarthi framed the problem the model is meant to solve.
Unlike conventional TTS systems that rely on text normalisation rules to determine pronunciation, Chariot's model uses contextual reasoning to decide how written expressions should be spoken.
Why this matters
Chariot's numbers are self-reported, and benchmarks against ElevenLabs, Cartesia, Microsoft, OpenAI, AWS, and Sarvam were run by Chariot itself, not an independent party. That's worth flagging before anyone builds production infrastructure around a 7.60% character error rate for Hindi. Still, the feature set is worth watching closely.
Code-mixing within a single sentence across any two of the 11 supported languages addresses something most Western-built TTS systems handle badly: how Indians actually speak, switching between Hindi, Tamil, or Bengali and English mid-thought. For developers building voice products for Indian markets, that's a more practical differentiator than a leaderboard position. Voice cloning from a few seconds of reference audio and a 1,000-character generation limit per request also tell you this is aimed at real deployment, not demos.
Founders should treat Chariot's claims as a starting point for their own testing, not a verdict. The real signal here isn't the model itself, it's that a DeepMind-and-FAIR-pedigreed team chose Indian languages as the problem worth solving first.
Related Reading
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
From Archives Kyndryl Study: Only 35% of Indian Firms Prepared for AI Cyberattacks
From Archives OpenAI Launches India API Inference Residency for Data Storage and Processing
Common Questions Answered
What languages does the Chariot-1-TTS-Multilingual model support?
Chariot-1-TTS-Multilingual supports 11 Indian languages plus English, with the ability to handle mid-sentence code-mixing between any two of these languages. This feature addresses a significant gap in Western-built TTS systems that typically struggle with code-mixing within a single sentence.
How does Chariot's contextual reasoning approach differ from conventional TTS systems?
Unlike conventional TTS systems that rely on text normalisation rules to determine pronunciation, Chariot's model uses contextual reasoning to decide how written expressions should be spoken. This approach allows for more natural and accurate pronunciation handling across different linguistic contexts.
What voice cloning capabilities does Chariot-1-TTS offer?
Chariot-1-TTS-Multilingual supports more than 50 voices and can clone a new voice from just a few seconds of reference audio. The model can generate speech from text blocks up to 1,000 characters long in a single request, making it versatile for various applications.
What should users know about Chariot's reported performance metrics?
Chariot's reported 7.60% character error rate for Hindi and all benchmarks against competitors like ElevenLabs, Cartesia, Microsoft, and AWS were self-reported and run by Chariot itself rather than by an independent third party. Users should consider this when evaluating whether to build production infrastructure around these performance claims.
Who founded Chariot and what is their background?
Chariot's founders previously worked at Google DeepMind and Meta's FAIR lab, bringing significant expertise in AI research to the development of this multilingual TTS model. This background demonstrates the team's deep experience in building advanced AI systems.
Further Reading
- Papers with Code - Latest NLP Research - Papers with Code
- Hugging Face Daily Papers - Hugging Face
- ArXiv CS.CL (Computation and Language) - ArXiv
