Chariot launches real-time speech AI model outperforming Google across Indian languages

Indian artificial intelligence startup Chariot has launched a real-time speech recognition model that it claims outperforms global and domestic tech competitors across 10 widely spoken Indian languages. The model, named Chariot-1-ASR Realtime, is built to transcribe speech instantaneously while matching or exceeding the accuracy of systems that require a complete recording before processing.

The development marks a notable milestone for India's domestic deep-tech ecosystem. Founded by former researchers from Google DeepMind and Meta's Fundamental AI Research lab, Chariot was among the first four teams selected under the Union government's IndiaAI Mission, which aims to build national sovereignty in foundational computing models and artificial intelligence infrastructure.

Performance benchmark against competing systems

According to technical evaluation data released by Chariot, the new model logged significantly lower transcription error rates across 79 public test sets compared with established market offerings.

In tests across all 10 supported Indian languages, Chariot-1-ASR Realtime delivered the lowest error rate among real-time transcription services, outperforming services from ElevenLabs, Deepgram and Sarvam. The model also took top position on three specialized tests designed for challenging conditions: the Monsoon benchmark for Indian-English and Hindi telephone calls, the Vaani benchmark for district-level Hindi regional accents, and the Kathbath-Hard benchmark for noisy environments.

Solving the speed versus accuracy tradeoff

Speech recognition tools typically force developers to choose between two operational methods. Real-time models generate text line by line as a speaker talks, making them essential for automated voice agents, customer support helplines and live broadcasting, though often at the expense of accuracy. Batch systems wait until an entire audio file is completed before generating text, allowing the algorithm to analyze full context to reduce mistakes.

Chariot said its architecture eliminates that compromise by training a broad foundational speech model before refining it for recognition tasks.

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

Commercial deployments and language capabilities

The model automatically detects regional accents and code-mixed speech, where speakers alternate between English and regional languages, without requiring the user to select a target language in advance. For Hindi transcriptions, the system can output English words in either Devanagari or Latin script depending on client requirements.

Chariot is targeting enterprise clients across regulated and high-volume sectors, including commercial banks, insurance firms, telecommunications operators, contact centres and public service helplines. The company plans to expand the underlying Chariot-1 foundation architecture into a single multimodal system capable of voice detection, language identification, speech generation and synthesis.