Skip to content
MukhyamantriMukhyamantri
Technology

12 languages, one sentence: an Indian voice model claims the top spot

IndiaAI-backed lab Chariot has launched a multilingual speech model covering 11 Indic languages and English, with more than 50 voices and voice cloning from a few seconds of audio. In the company's own tests it beat ElevenLabs and Sarvam on Hindi and Indic error rates.

· · 1 min read

India: Chariot, an Indian AI research lab backed by the IndiaAI Mission, has launched a speech generation model that reads a sentence for meaning before it reads it aloud, and it covers 11 Indic languages plus English.

The model, Chariot-1-TTS-Multilingual, offers more than 50 voices and can switch between two supported languages inside a single sentence. It can clone a voice from a few seconds of reference audio. Chariot says the design target is the messiness of real Indian text: code-mixing, numbers, abbreviations and mixed-language inputs that break most speech systems.

The company put its own numbers on the table. In Hindi intelligibility tests it recorded a character error rate of 7.60%, against 8% for ElevenLabs v3 and 8.21% for Sarvam Bulbul v3. Across nine other Indic languages it averaged 6.47%, against 6.87% for Sarvam Bulbul v3 and 8.38% for ElevenLabs v3. On an English verbalisation benchmark of more than 1,000 sentences across over 40 categories, including currency amounts, dates, website addresses, chemical formulas and stock tickers, it reported 93.78% accuracy. On NISQA, an automated audio quality measure, it scored 4.901 against 4.896 for ElevenLabs v3.

Chariot said the results come from its own evaluations and that the competing model versions and coverage varied across benchmarks. Independent testing has not been published.

The lab is one of the first four teams selected under the IndiaAI Mission, the government programme to build foundation models in India.

Speech is where Indian language AI gets tested hardest. A voice agent reading out a loan amount in Marathi, a railway announcement in Bengali, a customer helpline that must say a policy number correctly: each of these fails loudly when the model misreads a number or drops a syllable.

Character error rate counts how many characters a system gets wrong, and in speech it is the difference between a helpline saying ₹2,45,000 and ₹2,450. That is why the Indic numbers carry weight. These systems are being built for loan recovery calls, railway announcements and government advisories, where a misread digit lands on a real person.

Chariot has not disclosed pricing or when the model will be generally available.