NatWest and Nvidia Cut Misheard Words by 20% as AI Accent Recognition Rolls Out to RBS Customers
NatWest and Nvidia Cut Misheard Words by 20% as AI Accent Recognition Rolls Out to RBS Customers #
NatWest Group has announced a collaboration with Nvidia to improve speech recognition in its Spending Insights tool. Joint testing showed a 20% reduction in misheard words after the system was tuned to handle a wider range of UK accents and everyday financial vocabulary.
The project, disclosed on 14 October, uses Nvidia’s Nemotron open-source models, which engineers and researchers from both companies applied to real-world banking problems. The aim was to improve how the AI parses regional accents, colloquialisms, and financial terms rather than responding reliably only to standardised speech.
Spending Insights is NatWest’s generative audio-visual tool that lets customers explore their finances through voice and text conversations. It handles follow-up questions, interruptions, and shifts in topic, converting transaction records into summaries and charts. The bank says the collaboration combined Nvidia’s speech AI technology with NatWest’s customer data to tailor the system to UK retail banking.
The updated tool is being piloted with 75% of Royal Bank of Scotland customers, the group’s Scottish retail brand, before a planned broader rollout across NatWest Group. NatWest’s chief AI research officer said that if customers choose to speak to AI, the technology must be able to understand “the different accents, words and everyday language used across the UK.”
The partnership follows NatWest’s announcement of the Spending Insights tool in late September, part of a wider set of generative AI features that included an upgraded fraud triage agent in its Cora virtual assistant. The Nvidia work addresses a limitation in that launch: ensuring the voice interface works reliably for customers regardless of where in the UK they live or how they speak.
For Nvidia, the arrangement adds to its UK financial services work as the chipmaker looks to show practical uses for its Nemotron model family in domain-specific settings.
Copy link