Executives highlight the model’s focus on domain-specific accuracy and the real-language realities of Indian financial conversations.

Blue Machines AI launches Aurora, a multilingual speech-to-text model tailored for BFSI conversations in India, capable of handling Indian English, Hindi, Hinglish, and other multilingual speech—even over noisy or low-bandwidth connections.

Aurora targets language switching, multilingual code-mixing, noisy telephony audio, and BFSI-specific vocabulary such as EMIs, policy numbers, and transaction IDs to improve workflow accuracy.

Nirmit Parikh explains that India’s financial conversations are multilingual and non-scripted, and Aurora aims to reduce misunderstandings in EMI amounts, policy numbers, and repayment commitments to boost customer outcomes.

Aurora can be custom-trained on enterprise data to learn institution-specific names, terminology, geographies, accents, and interaction patterns, potentially reducing recognition errors by 40–45% on customized datasets.

With customer-authorised data, Aurora aligns more closely with an institution’s terminology and products, and retraining yields a 40–45% relative reduction in recognition errors on institutional datasets.

The model supports fine-tuning on enterprise data to learn institution-specific terminology and interaction patterns, driving a significant drop in errors on organization-specific material.

Aurora is trained to understand BFSI-specific vocabulary and financial identifiers used in workflows, such as EMIs, premiums, KYC, SIPs, policy numbers, and transaction IDs.

Throughput tests show Aurora handling up to 960 concurrent real-time streams at 320 ms latency and 2,400 streams at 1.12 seconds latency on Nvidia H100 GPUs.

Training and test data cover banking, lending, insurance, collections, and customer servicing with regional pronunciations, background noise, and telephony audio to mirror real BFSI interactions.

The company emphasizes sovereignty and control over data and customer interactions in deploying AI for Indian financial institutions.

Internal benchmarks report Semantic Word Error Rates of 1.51% (English), 2.43% (Hindi BFSI), and 5.52% (multilingual), with a BFSI Entity Error Rate of 4.23% for monetary amounts, policy numbers, and IDs.