India’s Sarvam AI Surpasses Global Rivals In OCR & Speech Models

Bengaluru-based artificial intelligence startup Sarvam AI has announced a major performance milestone.

The company stated that its latest vision and speech models have outperformed global technology leaders, including Google Gemini and ChatGPT, on key optical character recognition and text-to-speech benchmarks, particularly for Indian languages.

Sarvam AI co-founder Pratyush Kumar shared the results on the social media platform X. He said Sarvam Vision achieved a state-of-the-art accuracy of 84.3 per cent on the olmOCR-Bench English-only subset.

The model surpassed frontier systems such as Gemini 3 Pro and newer solutions like DeepSeek OCR 2.

On the OmniDocBench v1.5 English-only subset, Sarvam Vision recorded an overall score of 93.28 per cent.

Kumar said the model demonstrated strong capability in reading complex mathematical formulas and interpreting intricate document layouts. He added that the performance places Sarvam Vision close to the current global state of the art.

Kumar also highlighted advances in speech synthesis. Sarvam AI’s Bulbul V3 text-to-speech model supports 35 distinct voices. It covers all 22 scheduled Indian languages and handles documents with varying scan quality and diverse content types.

Kumar claimed Sarvam Vision currently leads performance benchmarks for Indian languages.

The model supports all constitutionally recognised Indian languages, making it one of the most comprehensive language platforms developed domestically.

The Vision series includes a 3-billion-parameter state-space architecture. It performs multiple tasks, including image captioning, scene text recognition, chart interpretation, and advanced table parsing.

These capabilities, however, allow accurate extraction of technical terminology from complex tables with merged rows and columns.

Kumar shared demonstrations showing Sarvam Vision extracting structured data from charts in the latest Economic Survey.

Additional examples highlighted its ability to understand and describe natural scenes from photographs with high contextual accuracy.

Sarvam AI said its broader objective is to democratise artificial intelligence across India.

The company, however, stated that it aims to help the country adopt transformative technologies with confidence and autonomy. It plans to build foundational AI components tailored to India’s linguistic and operational needs.

Union Minister for Information Technology Ashwini Vaishnaw acknowledged the achievement in a recent post on X. He said Sarvam AI’s progress reflects the growing impact of India’s national AI mission and its push for homegrown innovation.