Originals · Startup · Startups & Fintech
IndiaAI backed Chariot launches a text to speech model that reads context, not just words, across 11 Indian languages and English
The Delhi speech lab, formerly Gan.ai, says Chariot-1-TTS-Multilingual tops rival systems from ElevenLabs, OpenAI, AWS and Sarvam on its own benchmarks; independent tests are still to come.
Chariot, a Delhi headquartered AI research lab that was previously called Gan.ai, launched Chariot-1-TTS-Multilingual on November 12, 2026. The text to speech model covers 11 Indian languages and English, offers more than 50 voices, can switch between any two supported languages inside a single sentence, and can clone a voice from a few seconds of reference audio. A single request can carry up to 1,000 characters, according to Analytics India Magazine and CIOL.
The pitch is that the model understands what it is reading before it speaks. Conventional speech systems lean on hand written normalisation rules to decide how to say numbers, dates, currency amounts and abbreviations. Chariot says its model makes those calls itself, using context: Analytics India Magazine gives the example of telling apart B+ as a blood group and B+ as an exam grade. "People use numbers, symbols, and abbreviations, and move between languages in the same sentence," co-founder Parth Sarthi said in a statement. He told CIOL the model is built on the same Chariot-1 foundation as the company's speech recognition, which it launched earlier as Chariot-1-ASR Realtime for 10 Indian languages and Indian English.
The performance claims are the company's own. In an evaluation dated November 11, Chariot reports a character error rate of 7.60% on a Hindi suite of 644 prompts, the lowest average error rate across a nine language Indic suite (Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu), 93.78% accuracy on an English test of 1,029 sentences that covers currency, dates, URLs and stock tickers, and a NISQA speech quality score of 4.901. Moneycontrol reports the comparison set included models from ElevenLabs, Cartesia, Microsoft, OpenAI, AWS and Sarvam. Analytics India Magazine cautions that these are company reported results and that benchmark coverage varies across the competing models.
Chariot was founded by Parth Sarthi and Suvrat Bhooshan, who previously worked at Google DeepMind and Meta's FAIR lab, and it is one of the teams chosen to build foundation models under the government's IndiaAI Mission; Moneycontrol describes it as one of the first four teams selected. The company says the model is aimed at voice agents, customer service, education platforms and multilingual public helplines. Its website showcases earlier voice work for Zomato, Doordarshan, Coca-Cola, Cadbury and HDFC Bank.
Why it matters: India's voice AI race is now as much about getting a rupee amount, a date or a code mixed Hinglish sentence right as about sounding natural. The thing to watch is whether Chariot's numbers hold up when developers and independent evaluators test the model against Sarvam and the global players on their own data.
