Sarvam AI updates model to boost Indian language OCR accuracy

BENGALURU: Sarvam AI has released an upgraded model to read handwritten records, complex forms and multi-page tables in Indian languages, as the Bengaluru-based startup seeks wider business use for its document-processing tools.Sarvam Vision 2.1, announced on Thursday, can convert scanned documents into digital text and extract specific information from forms and tables. It works across English and 22 Indian languages, the company said.The earlier version, introduced in February, struggled with some complex documents and could occasionally generate text that was not in the original, according to Sarvam. Users had also raised concerns about the cost of running it at scale. The company said it trained the update on a mix of real and artificially generated documents and made changes to lower running costs. It did not quantify the price reduction.Alongside the model, Sarvam released a test set of 6,909 document samples to measure how accurately AI systems read different scripts. Of these, 6,609 are in Indian languages. On this benchmark, Vision 2.1 scored 87.3% overall, ahead of Bodhan Indic-OCR’s 84.9% in Sarvam’s comparison. The samples include newspapers, textbooks and historical material dating back to 1800.The results also show gaps. Vision 2.1 scored 53.9% on Santali and 54.8% on Kashmiri, against more than 90% on Hindi, Kannada and Telugu. Bodhan Indic-OCR outperformed it on Santali in Sarvam’s tests.On a separate international test measuring how well models preserve document structure, Sarvam ranked second behind PaddleOCR. The company acknowledged that strong benchmark results do not necessarily establish how useful a model will be on real-world documents. Its performance comparisons are based on Sarvam’s own evaluations.You use AI every day. Now get your AI Quotient. Take the AIQ test.