What is the story about?
Bengaluru-based startup Sarvam AI made headlines this week with the launch of two powerful Artificial Intelligence tools known as Sarvam Vision and Bulbul V3, claiming to outperform international AI chatbots Google Gemini and OpenAI's ChatGPT.
The standout, Sarvam Vision, is specifically designed for document reading and text recognition. OCR technology allows machines to read text from images, scanned documents, or handwritten notes.
Sarvam Vision
Beyond simple text extraction, it can extract data points from graphs, interpret trends from charts, and preserve complex table structures, even when tables are nested or visually complicated. The model is also accurate in 22 languages, including major Indian languages like Hindi, Bengali, Tamil, Telugu, and Marathi.
Sarvam Vision vs Gemini Pro 3 and ChatGPT
According to Pratyush Kumar, co-founder of Sarvam AI, Sarvam Vision achieved an impressive 84.3% accuracy on the olmOCR-Bench, surpassing other leading OCR tools such as Gemini 3 Pro at 80.20%, DeepSeek OCR v2 at 78.80%, and ChatGPT at 69.80%.
The model also scored 93.28% on OmniDocBench v1.5, which is basically a benchmark used to measure how well AI models can read and understand documents in various formats, such as columns, tables, and mixed-layout content. On this benchmark, Sarvam Vision outperformed Gemini 3 Pro at 91.6% and ChatGPT 5.2 at 86.56 %t.
In addition, Sarvam Vision topped the word accuracy test with 87.36%, compared to Gemini 3 Pro at 82.51% and ChatGPT 5.2 at 38.60%.
Bulbul V3
Alongside OCR, Sarvam AI has also launched Bulbul V3, a next-generation text-to-speech tool designed to deliver natural, expressive and production-ready voices for Indian languages. The model offers over 35 high-quality voices across 11 Indian languages which sound natural and expressive.
"People switch languages mid-sentence. Accents vary by region. Names, abbreviations, and emotions matter as much as words. To work in India, voice has to handle all of this without breaking," the blog post read.
Bulbul V3 has been tested in independent listening studies, and it scored high on three important points. First, it sounds very natural and human-like, whether in high-quality audio or standard phone calls. Second, it is robust, meaning it can read tricky text such as code-mixing and numerics with low error rates and third, it is stable, as it can perform well even in long recordings or high-volume use.
The team realised that if a voice doesn't feel right in the first few seconds, listeners quickly lose interest. So, they built Bulbul V3 to focus on pacing, emphasis, and emotional tone, not just reading words aloud. It figures out where to emphasize, when to pause, and how to adjust tone and speed.
The standout, Sarvam Vision, is specifically designed for document reading and text recognition. OCR technology allows machines to read text from images, scanned documents, or handwritten notes.
Sarvam Vision
Beyond simple text extraction, it can extract data points from graphs, interpret trends from charts, and preserve complex table structures, even when tables are nested or visually complicated. The model is also accurate in 22 languages, including major Indian languages like Hindi, Bengali, Tamil, Telugu, and Marathi.
Sarvam Vision vs Gemini Pro 3 and ChatGPT
According to Pratyush Kumar, co-founder of Sarvam AI, Sarvam Vision achieved an impressive 84.3% accuracy on the olmOCR-Bench, surpassing other leading OCR tools such as Gemini 3 Pro at 80.20%, DeepSeek OCR v2 at 78.80%, and ChatGPT at 69.80%.
The model also scored 93.28% on OmniDocBench v1.5, which is basically a benchmark used to measure how well AI models can read and understand documents in various formats, such as columns, tables, and mixed-layout content. On this benchmark, Sarvam Vision outperformed Gemini 3 Pro at 91.6% and ChatGPT 5.2 at 86.56 %t.
In addition, Sarvam Vision topped the word accuracy test with 87.36%, compared to Gemini 3 Pro at 82.51% and ChatGPT 5.2 at 38.60%.
Bulbul V3
Alongside OCR, Sarvam AI has also launched Bulbul V3, a next-generation text-to-speech tool designed to deliver natural, expressive and production-ready voices for Indian languages. The model offers over 35 high-quality voices across 11 Indian languages which sound natural and expressive.
"People switch languages mid-sentence. Accents vary by region. Names, abbreviations, and emotions matter as much as words. To work in India, voice has to handle all of this without breaking," the blog post read.
Bulbul V3 has been tested in independent listening studies, and it scored high on three important points. First, it sounds very natural and human-like, whether in high-quality audio or standard phone calls. Second, it is robust, meaning it can read tricky text such as code-mixing and numerics with low error rates and third, it is stable, as it can perform well even in long recordings or high-volume use.
The team realised that if a voice doesn't feel right in the first few seconds, listeners quickly lose interest. So, they built Bulbul V3 to focus on pacing, emphasis, and emotional tone, not just reading words aloud. It figures out where to emphasize, when to pause, and how to adjust tone and speed.
