Blue Machines AI Launches Aurora Speech Model for Indian BFSI

Bengaluru-based voice AI orchestration startup Blue Machines AI has launched Aurora, a speech-to-text model built for banking, financial services and insurance (BFSI) conversations in India. The company says Aurora recorded a 1.51% Semantic Word Error Rate (WER) on English BFSI conversations, 2.43% on Hindi BFSI speech, and 5.52% across multilingual speech in its internal benchmarks.

Aurora also recorded a BFSI Entity Error Rate of 4.23% for financial information such as monetary amounts, interest rates, policy numbers, account references, and transaction IDs. The company says the model recorded a P50 latency of 236 milliseconds.

P50 latency measures the median response time, indicating how quickly 50% of requests finish faster than the other 50%.

The model handles real-time conversations where customers switch between Indian languages and English financial terms. 

Its training and evaluation datasets cover banking, lending, insurance, collections, and customer servicing, including Indian English, Hindi, Hinglish, multilingual, and code-mixed speech.

Blue Machines AI said Aurora was evaluated against leading speech-to-text models using the same audio inputs and scoring methodology. It claimed the model delivers higher accuracy at lower latency in its internal evaluation and is designed for streaming financial conversations.

“Aurora reflects our commitment to building sovereign AI infrastructure for Indian enterprises,” said Nirmit Parikh, Founder and CEO of Blue Machines AI. “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.”

Aurora is trained to recognise BFSI terms including EMIs, outstanding amounts, foreclosure charges, disbursals, KYC, premiums, SIPs, NAVs, policy numbers, transaction IDs, and payment dates. It is also designed to handle numbers, currencies, percentages, and financial identifiers used in financial workflows.

“Aurora’s cache-aware FastConformer encoder and streaming transducer decoder enable it to retain context while processing speech incrementally,” said Abhishek Ranjan, Chief Technology Officer at Blue Machines AI.

Ranjan said the model is also optimised for multilingual and code-mixed speech, low-latency inference, and high-concurrency environments. In internal throughput tests, Aurora supported 960 concurrent real-time streams per H100 at a 320-millisecond operating point and 2,400 concurrent streams per H100 at a 1.12-second operating point.

Blue Machines AI has built training pipelines that allow Aurora to be adapted using customer-authorised enterprise data. The company claimed customer-specific retraining can reduce recognition errors by 40–45% relative to the base model on institution-specific datasets. The adaptation can account for proprietary product names, terminology, geographies, accents, and interaction patterns.

Aurora integrates with Blue Machines AI’s enterprise CX AI platform and can be used across customer acquisition, onboarding, lending, collections, servicing, insurance, claims, and customer support.

The model can be deployed through a managed cloud, an enterprise VPC, or on-premises infrastructure, allowing financial institutions to choose deployment options based on their security, data residency, and governance requirements.

Blue Machines AI recently partnered with Muthoot Finance to deploy conversational AI for gold-loan lead engagement and qualification. The deployment has reached 1.39 million leads, with 63.1% customer connectivity and a 150% increase in branch-visit conversion. It has also cut the average journey completion time by more than 70%, from over seven minutes to around two minutes.