New Delhi, Oct 22: Chariot, a frontier AI research lab founded by former researchers from Google DeepMind and Meta's FAIR lab, and one of the first four teams selected under the IndiaAI Mission, today launched Chariot-1-ASR Realtime.
Most speech recognition models are only built for that one task. Chariot-1, however, is a foundation model for speech. It learns speech the way large language models learn text: broadly and once, and is then specialised for specific tasks. Speech recognition is one of them.
Speech recognition comes in two kinds. Live systems, used in voice agents, call centres and captions, write text as someone speaks. Batch systems wait for a recording to finish and, with the whole recording to work from, are usually more accurate. Chariot-1-ASR Realtime is a live system, and it beats the best of both.
Across 79 public test sets, it made 9% fewer errors than the strongest batch system, Sarvam's Saaras v4, and 26% fewer than Google's Gemini 3.5 Transcribe. Against real-time services from Sarvam, ElevenLabs, Deepgram and others, it had the lowest error rate in all 10 languages. In Hindi, it made 23% fewer errors than the next best.
It inherits Chariot-1’s broad understanding of speech, and that shows most on real-world audio:
- First on real Hindi and Indian-English phone calls
- First on Hindi accents from districts across India
- First on noisy recordings
Parth Sarthi, Co-Founder, Chariot, said:
"Speech recognition has always made you choose: fast, or accurate. For Indian languages, Chariot-1-ASR Realtime ends that choice. We didn't measure it against one rival, or against an average. In every language, we put it up against the strongest system we could find, the best of the best, including systems that get to hear the whole recording first, and it came out ahead in all ten. This is what frontier AI built in India looks like."
For businesses, the live transcript is now also the most accurate one. The same model powers voice agents in real time and produces the transcripts that compliance and analytics teams rely on. It handles regional accents and code mixing without being told the language. It can write English words in Latin script, or in Devanagari for Hindi. Banks, insurers, telecom operators, contact centres and public helplines can serve customers across 10 of India's most widely spoken languages with one model.
Chariot’s broader approach treats speech understanding and speech generation as connected capabilities, not separate tasks. Voice products today stitch together separate models for detecting speech, identifying the language, transcribing and speaking back, and each one relearns the same things about speech. Chariot sees these as one problem, speech reasoning, and is building Chariot-1 as a single model to solve it, from voice detection and language identification to transcription, speech generation and even singing. Chariot-1-ASR Realtime is the first speech recognition release from the Chariot-1 model family, with further developments expected across speech recognition and speech generation.
Suvrat Bhooshan, Co-Founder, Chariot, said:
"We believe speech is one problem, not many. Understanding a sentence and saying one rely on the same knowledge, so we built a single model to learn it once. Chariot-1-ASR Realtime proves our hypothesis: the same foundation, trained further for speech recognition, is now ahead of every system we tested on Indian languages. Because the model learns how Indian languages sound, how people code-mix them and how accents shape them, every gain in that foundation shows up in speech recognition, often in places we never aimed for. This is the first speech recognition model from Chariot-1, and the same foundation will power what we release next."
Chariot-1-ASR Realtime is available on the Chariot platform at platform.chariot.in, alongside Chariot's text-to-speech capability.
