Chariot releases 12-language speech model designed to read numbers in context

Chariot has released a speech-generation model designed to distinguish a blood group from a school grade—even when both are written as B+. Launched on November 12, 2026, Chariot-1-TTS-Multilingual supports 11 Indic languages and English. Its pitch is less work preparing text for speech, with the model itself deciding how numbers, symbols and abbreviations should sound.

The same symbol, two spoken meanings

Text-to-speech software must turn written shorthand into spoken words. Chariot says its model makes those decisions internally, reducing the need for custom text-normalisation rules or separate prompts that prepare an input before it is read aloud. The distinction is not simply how a voice pronounces a word, but which spoken expression preserves the intended meaning.

The company supplies illustrative examples of the decisions it wants the model to make. These are examples authored by Chariot, rather than independently verified demonstrations:

• Blood group B+ should become “blood group B positive,” while a grade of B+ should become “a grade of B plus.”
• ₹1,250 should be read as “one thousand two hundred and fifty rupees,” preserving both the amount and its currency.
• Chapter IV should become “chapter four,” treating the Roman numeral as a chapter number rather than reading its letters.

Language switching is another part of the release. Chariot says a sentence can combine any two supported languages, rather than requiring the entire input to stay in one language. The model also offers more than 50 voices and can create a cloned voice from a few seconds of reference audio. Each generation accepts up to 1,000 characters.

Three tests, different competitive gaps

Chariot separates its evaluation into three jobs: preserving the intended words, choosing an acceptable spoken interpretation and producing audio that scores well with an automated quality evaluator. It says it ranked first on those measures among the configurations compared in its November 11 evaluation. That is a claim about selected model versions, not every competing speech system.

For Hindi intelligibility, the company generated speech, transcribed it with speech-recognition systems and compared the transcripts with normalised reference text. Lower character error rates mean closer matches. Its Hindi suite contained 644 configured prompts spanning general text, numerical and technical challenges, and native conversational material. Hindi scores average results from three speech recognisers.

Across nine other Indic languages, Chariot reported a mean character error rate of 6.47%, against 6.87% for Sarvam Bulbul v3 and 8.38% for ElevenLabs v3. That comparison includes models with complete coverage and weights 27 language-and-dataset scores equally. Chariot also says it had the lowest average error rate in eight of the nine individual languages.

The English interpretation test produced a narrower lead: 93.78% accuracy for Chariot versus 93.51% for ElevenLabs v3. It covered 1,029 sentences in 42 categories, including dates, currencies, website addresses, chemical formulas and stock tickers. The measure asks whether a judged output gives an acceptable spoken reading—not whether a listener prefers the voice.

Audio quality was assessed separately using NISQA, an automated predictor, on 1,088 plain-English prompts. Chariot scored 4.901, compared with 4.896 for ElevenLabs v3. These are company-run results, with rival versions and coverage varying by benchmark. Moneycontrol notes that the results have not been independently verified.

Language support is wider than the test suite

The supported languages are Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu, Urdu and English. Chariot describes its evaluation suite as 8,655 configured text cases across English and ten Indic languages. Its nine-language intelligibility comparison covers the listed Indic languages other than Hindi and Urdu; Hindi receives its own comparison.

Chariot is targeting voice agents, customer service, educational platforms and multilingual public helplines, according to Analytics India Magazine. Those planned uses combine the release’s two main jobs: speaking across languages and reading written details appropriately. Developers can choose voices from the catalogue or create custom ones using reference recordings.

The launch announcement offers a text-to-speech trial. Chariot recommends starting with a mixed-language sentence or an amount, date or abbreviation that an application actually needs to say aloud.