NatWest and NVIDIA have adapted speech AI to UK accents and banking language, cutting misheard words by 20% in NatWest’s Spending Insights tool.
UK-based NatWest Group has worked with NVIDIA to adjust automatic speech recognition (ASR) to the accents, pronunciations, and colloquial financial language used by its customers across the UK. The work supports Spending Insights, a conversational AI service developed by the bank that lets customers query their finances by voice and text. According to the announcement, the adapted system misheard 20% fewer words and recorded a lower overall Word Error Rate, the standard measure commonly used to assess transcription accuracy.
Access to Spending Insights is being extended in stages. 75% of Royal Bank of Scotland customers can currently use it, with availability across the wider NatWest Group to follow. Users can ask about their spending, explore trends, and receive tailored insights, and they can interrupt, add context, or ask follow-up questions along the way. Answers combine conversational responses with interactive charts and visualisations.
How the speech model was adapted
The service works in two stages. NVIDIA Nemotron ASR, drawn from NVIDIA’s open-source Nemotron model family, first turns the customer’s speech into text. NatWest’s in-house Small Language Models then read that transcript, work out what is being asked, and formulate a reply.
The project began with testing involving 90 NatWest employees from different parts of the UK. Their accents included Glaswegian, Northern Irish, Welsh, Yorkshire, Mancunian, Midlands, and Southern English, as well as non-British accents. Participants recorded around 3,000 banking questions, both reading scripted versions aloud and rephrasing them in their own words to capture informal phrasing. Researchers from both companies then mapped where the models performed well and where specific accents, words, or constructions caused difficulties.
Testing showed how minor transcription errors can change the meaning of a request. Merchant names, dates, and financial terminology proved particularly difficult, despite being central to interpreting customer queries correctly.
NatWest then extended the work using anonymised contact-centre transcripts, exposing the model to a broader range of real-world banking language, and further improving its recognition of merchant names and financial terms.
The transcripts were anonymised and kept within NatWest’s environment. No customer data was shared with NVIDIA, and any model adaptations made using NatWest data are ringfenced to the bank.
Accuracy as a prerequisite for voice banking
For banks introducing voice interfaces, the project highlights a practical constraint: speech recognition must handle regional speech and domain-specific vocabulary in real time before downstream language models can respond reliably. Maja Pantic, Chief AI Research Officer at NatWest Group, noted that customers increasingly use AI in their daily lives and that, if they choose to speak to it, the technology needs to understand the accents and everyday language used across the UK. Aser Blanco, Global Head of Banking at NVIDIA, said real-time operation and varied terminology make speech recognition considerably harder, and that the collaboration allowed NVIDIA to apply Nemotron models to real-world banking use cases and identify where they could be improved.
The two companies plan to continue refining the speech-recognition technology as conversational banking services develop.
