Bengaluru: Homegrown AI lab Chariot launched Chariot-1-TTS-Multilingual on Thursday, a text-to-speech model that spans 11 Indic languages and English, carries more than 50 voices and can mix any two supported languages inside a single sentence.
The harder problem in speech is knowing how a written string should sound. A payment amount needs its spoken value and currency. "B+" reads one way as a blood group and another as a grade. Dates, abbreviations and technical expressions throw up the same choice. Chariot says that reasoning is built into the model, cutting the need for hand-written normalisation rules or engineered prompts.
The company's evaluation compares selected versions from ElevenLabs, Cartesia, Microsoft, OpenAI, AWS and Sarvam, with coverage varying by benchmark. Chariot says its model ranked first on three measures.
On intelligibility it reports the lowest character error rate, 7.60 per cent on Hindi and 6.47 per cent across nine other Indic languages, tested on general text, numerical and technical stress tests and native conversational text. On reasoning accuracy it reports the highest English verbalisation accuracy at 93.78 per cent, on a benchmark of more than 1,000 sentences across over 40 categories, including currency amounts, dates, ordinals, URLs, chemical formulas and stock tickers. On quality it claims the highest plain-English NISQA score, an automated prediction of speech quality.
The model can clone a voice from a few seconds of reference audio and generate speech from up to 1,000 characters in one request. All of these figures are the company's own; they have not been independently verified.
Voice cloning cuts both ways. Consent controls, deepfake detection and fraud prevention will decide how quickly this class of model moves into call centres, classrooms and public services. For Indian languages, the contest has moved to how the machine sounds.