Chariot Claims Its Speech AI Beats Google in 10 Indian Languages

01.A live transcript in ten Indian languages

Chariot, a Delhi-based AI research lab, on Thursday launched Chariot-1-ASR Realtime, a speech recognition model that transcribes speech as it is being spoken. The company says it covers ten of India's most widely spoken languages, Hindi, Marathi, Bengali, Tamil, Telugu, Kannada, Malayalam, Gujarati, Punjabi and Odia, along with Indian English, and that it is live on the Chariot platform now. The launch was reported by Moneycontrol and by ET EnterpriseAI.

The model is built on Chariot-1, the company's foundation speech model. Chariot says it can run in two modes: a balanced mode that returns text sooner, meant for voice agents and live captions, and a high-accuracy mode that waits for more context before committing to text, meant for call and meeting transcripts. It can also write borrowed English words in Latin script or in the script of the language being spoken.

02.Why real-time is the harder problem

Speech recognition usually forces a choice. Batch systems hear the whole recording before writing a word, which lets them use the end of a sentence to settle a word at the start. Streaming systems have to commit as the audio arrives, and that normally costs accuracy. That is why the most accurate transcription has traditionally been the slow kind.

Chariot's pitch is that its model does not pay that cost. In the company's published evaluation, the streaming model posted the lowest word error rate, the share of words a system gets wrong, in every language it tested, against both streaming services and batch models. If that holds up under independent testing, it would matter for anything that has to react to speech in real time. For call centres and voice agents, a delay of a second or two can be the difference between a natural conversation and a frustrating one.

03.The numbers Chariot is claiming

Chariot says it tested the model on 79 public test sets across ten Indian languages, about 153,000 utterances covering read, spontaneous, conversational, phone and noisy audio, and scored every system with the same code. Across those sets it reports 9 per cent fewer errors than Sarvam's Saaras v4, which it calls the strongest batch system it compared against, and 26 per cent fewer than Google's Gemini 3.5 Transcribe. In Hindi, it says it made 23 per cent fewer errors than the next-best system.

The company also says the model ranked first on three benchmarks built from real-world audio: Monsoon, for Hindi and Indian-English phone calls; Vaani, for Hindi accents from districts across the country; and Kathbath-Hard, for noisy recordings. These are the company's own results, published by Chariot, so they are claims to be tested rather than settled facts. The systems it compared against include real-time and batch services from Sarvam, ElevenLabs, Deepgram and others.

04.Who is behind it

Chariot is a frontier AI research lab headquartered in Delhi, founded by Parth Sarthi and Suvrat Bhooshan, with a team drawn from Google DeepMind, Meta's FAIR lab, and institutes including Stanford, Georgia Tech, IIT Delhi and IIT Bombay. According to ET EnterpriseAI, it was among the first four teams selected under the IndiaAI Mission to build the country's foundation models.

The company's idea is that audio is a modality, not a domain. Instead of bolting together separate models for detecting speech, identifying the language, removing noise, transcribing and punctuating, it is training one general speech model and then specialising it for each task. Chariot-1-ASR Realtime is the first speech-recognition release from that family. Moneycontrol reports the company is using NVIDIA's NeMo framework to develop an eight-billion-parameter model for real-time text-to-speech as well.

05.What it means for Indian users and businesses

The practical use case is voice in Indian languages at scale. Banks, insurers, telecom operators, contact centres and public helplines all deal with customers who speak different languages, often mixing English into the same sentence. A single system that handles ten languages, copes with regional accents and code-mixed speech, and does not need to be told the language in advance is a useful building block for those services.

For builders and students, the more interesting signal is the direction. An IndiaAI Mission-backed lab is competing on benchmarks with global players in the languages that matter here, and publishing its numbers for anyone to check. The caveat is that the headline comparisons come from Chariot's own evaluation. Until independent reviewers reproduce them, treat the ranking as a strong claim from a serious lab, not a settled scoreboard.

Moneycontrol also notes that the speech AI market is getting crowded, with competition pushing prices down across the industry, so a benchmark win will have to be paired with a business that can live with that pressure. Chariot's decision to publish its evaluation in detail, including the test sets and the competitors it ran against, at least gives other labs and customers something concrete to argue with.