Introducing Chariot-1-TTS-Multilingual

We’re releasing Chariot-1-TTS-Multilingual, a text-to-speech model for the way people actually write and speak. It can switch languages mid-sentence and use context to read prices, dates, abbreviations and symbols aloud.

Multilingual speech for real applications

• 11 Indic languages and English. Generate speech in Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu, Urdu and English.
• 50+ voices. Choose from the voice library for conversational interfaces, narration, learning and other voice applications.
• Code-mixing across all supported language pairs. Generate speech from sentences that combine any two supported languages.
• Instant voice cloning. Create a voice from a few seconds of reference audio and use it to generate speech.
• Up to 1,000 characters per generation. Generate a complete response or passage in one request.

How the model reads text in context

A written amount needs the right spoken value and currency. An abbreviation needs an appropriate expansion. And the same symbol can mean different things depending on what surrounds it.

Chariot-1-TTS-Multilingual makes these decisions in the model itself, reducing the need to maintain custom normalization rules or write separate prompts to prepare text for speech.

Three tiers of speech evaluation

We measured how faithfully the model conveys the intended words, how accurately it interprets written expressions, and how an automated evaluator scores the resulting audio. Chariot ranked first on the three metrics below among the configurations compared in our 11 November 2026 evaluation.

Tier 1 · Intelligibility

We transcribe the generated audio with automatic speech recognition and compare the transcript with a normalized reference. The resulting character error rate (CER) measures how faithfully the speech preserves the intended content.

On the Hindi suite, Chariot records 7.60% CER across 644 configured prompts, covering general, robustness and native conversational text.

The nine-language Indic suite covers Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu. General text includes assistant and commerce phrasing, robustness text adds numerical and technical expressions, and conversational text comes from IndicVoices-R.

Among models with complete coverage, Chariot has the lowest nine-language average CER on each dataset. It also has the lowest average CER in eight of nine individual languages when the three test sets receive equal weight.

Tier 2 · Reasoning accuracy

We define reasoning accuracy as verbalization accuracy, the share of judged outputs with an acceptable spoken interpretation. The English suite has 1,029 sentences across 42 categories, including currency, dates, ordinals, URLs, chemical formulas, stock tickers and codes. Chariot achieves 93.78%.

Tier 3 · Audio quality

We use NISQA, an automated speech-quality predictor, on 1,088 plain-English prompts from Seed-TTS-eval. Chariot records 4.901, the highest score among the compared configurations.

Evaluation methodology

The evaluation suite includes 8,655 configured text cases across English and ten Indic languages. Each table ranks the model versions included in that benchmark.

• Intelligibility. We compared transcripts of the generated audio with normalized references. The nine-language summary gives equal weight to each language and dataset.
• Verbalization accuracy. We judged whether the audio gave an acceptable spoken reading of the input. Accuracy is the share of resolved judgments marked correct.
• Audio quality. We used NISQA to estimate the quality of the generated English audio.

The scores come from Chariot’s TTS Evaluation Standard: Proposal. The benchmark datasets are linked below.