Blue Machines AI has launched Aurora, a multilingual speech-to-text model designed specifically for banking, financial services and insurance (BFSI) conversations in India.
The model is built to process real-time conversations involving Indian English, Hindi, Hinglish and other multilingual and code-mixed speech, including interactions over noisy or low-bandwidth telephone connections.
According to internal benchmarks conducted by Blue Machines AI on BFSI datasets, Aurora recorded a Semantic Word Error Rate (WER) of 1.51% for English, 2.43% for Hindi BFSI conversations and 5.52% across multilingual speech. It recorded a BFSI Entity Error Rate of 4.23% for information including monetary amounts, interest rates, policy numbers, account references and transaction IDs.
The company said the datasets used for testing covered banking, lending, insurance, collections and customer servicing, and included regional pronunciation patterns, background noise and telephony audio.
Aurora has been trained to recognise financial terminology and entities such as EMIs, outstanding amounts, foreclosure charges, disbursals, KYC, premiums, SIPs, NAVs, policy numbers and transaction IDs.
“India’s financial conversations do not happen in a single language or follow a standard script. When AI misunderstands an EMI amount, policy number or repayment commitment, it can change the customer outcome,” said Nirmit Parikh, Founder and CEO, Blue Machines AI. “By building Aurora in India, we are giving financial institutions speech intelligence designed for how their customers naturally communicate, while ensuring greater control over their data, models and customer interactions.”
Abhishek Ranjan, Chief Technology Officer at Blue Machines AI, said the model has been optimised for multilingual and code-mixed speech, low-latency inference and high-concurrency environments. “Crucially, it is evaluated on its ability to accurately recognise the entities that drive financial workflows, not merely the surrounding sentences,” he said.
According to Ranjan, internal throughput tests showed Aurora supporting 960 concurrent real-time streams per Nvidia H100 GPU at a 320 millisecond operating point and 2,400 concurrent streams per H100 at a 1.12-second operating point.
Blue Machines AI said Aurora can also be adapted using customer-authorised enterprise data to account for institution-specific product names, terminology, geographies, accents and interaction patterns. Its internal evaluations showed such retraining resulted in a 40-45% relative reduction in recognition errors compared with the base model on institution-specific datasets.
Aurora can be deployed through managed cloud infrastructure, within an enterprise virtual private cloud (VPC), or on-premises.