Chariot launches Chariot-1-ASR Realtime, a streaming speech recognition model for 10 Indian languages

Chariot, a frontier AI research lab, launched Chariot-1-ASR Realtime on October 22, a speech recognition model that converts speech to text while the speaker is still talking. The company claims the model outperformed every streaming and batch transcription system it tested across 10 Indian languages. The model is available on the Chariot platform. Analytics India Magazine reported the launch.

Languages and modes

The model supports Hindi, Marathi, Bengali, Tamil, Telugu, Kannada, Malayalam, Gujarati, Punjabi and Odia. It also supports Indian English.

It offers two modes:

• Balanced: produces text faster, aimed at voice assistants and live captions.
• High-accuracy: considers more audio before producing text, aimed at call transcripts, meeting notes and dictation.

The model can detect the language automatically or use a language selected by the application. It can write borrowed English words in either the English alphabet or the script of the language being spoken.

Chariot's benchmark claims

Chariot says it tested the model on 79 public test sets from 11 collections, covering around 153,000 audio recordings. The tests included everyday conversations, telephone calls, read-aloud speech and recordings with background noise. All results below come from Chariot's own tests.

Against systems that process an entire recording before producing text, Chariot reports 26% fewer errors than Google Gemini 3.5 Transcribe and 9% fewer than Sarvam's Saaras v4.

Against streaming systems, Chariot reports:

• 17% fewer errors than Sarvam Saaras v4
• 34% fewer errors than Google Gemini 3.5 Transcribe Live
• 57% fewer errors than ElevenLabs Scribe v2 Realtime

Chariot says its model was not told which language was being spoken during these tests, while competing systems were given the language settings that worked best for them.

Dataset-level results (lower is better):

• Monsoon, a public collection of unscripted telephone conversations: 5.04% error rate on Hindi calls, compared with 8.27% for OpenAI's gpt-transcribe.
• Monsoon, Indian English calls: a 3% error rate, which Chariot says matched results from Microsoft Azure Speech and Zoom Scribe v2 Pro on the Open ASR Leaderboard.
• Vaani, a collection of Hindi recordings from speakers across different districts: 17.21%, compared with 17.86% for Sarvam Saaras v4.

Model background

Chariot-1-ASR Realtime is based on Chariot-1, the company's broader speech model, adapted for transcription. Chariot says it plans to use the same underlying model for other speech tasks, including speech generation and language identification.

Chariot co-founder Parth Sarthi said: "Speech recognition has always made you choose: fast or accurate. For Indian languages, Chariot-1-ASR Realtime ends that choice."

Pricing not yet disclosed. Parameter count and context window were not given.

What this means

The headline claim is that one model covers both low-latency and high-accuracy use cases, with the choice exposed as two modes rather than two separate products. For teams building voice assistants, live captioning, call transcription or dictation in Indian languages, that means a single model could serve both kinds of workload.

The comparisons deserve careful reading. Every figure comes from Chariot's own testing. Chariot ran its model without language hints while competitors received their best-performing language settings, which Chariot presents as a harder test for its own model. The Vaani Hindi gap (17.21% against 17.86%) is narrow, while the Monsoon Hindi telephone result (5.04% against 8.27% for gpt-transcribe) is wider. Independent evaluation on a team's own audio, especially noisy telephone speech, would be needed to confirm how these results hold in production.

Chariot also states that Chariot-1 will underpin speech generation and language identification, so this release is the first transcription product on a broader speech model.