NatWest Group is working with NVIDIA to improve how artificial intelligence understands the way people across the UK talk about the economy. The collaboration has reduced the number of misheard words by 20% across the accents tested, supporting the bank’s push to make conversational banking more natural and accurate.
Announced on 14 October 2026, the work uses NVIDIA Nemotron automatic speech recognition technology alongside NatWest’s in-house AI research. The technology supports Spending Insights, a conversational banking tool that allows customers to explore their finances through voice and text, ask follow-up questions,, and view interactive spending information.
The development addresses a practical challenge for financial institutions adopting AI: a system may be capable of answering a question, but its response can still be wrong if it misunderstands what the customer said in the first place.
NatWest said Spending Insights was being made available to 75% of Royal Bank of Scotland customers as part of an ongoing rollout. That figure describes access during the rollout, not the proportion of customers actively using the service. The bank plans to extend the experience more widely across the group.
Why AI speech recognition matters for banking
Conversational banking promises to let customers manage their finances in everyday language rather than navigate multiple screens, menus,, and filters.
A customer might ask which purchases increased last month, how much they spent on groceries in the previous quarter,, or which subscriptions appear regularly in their transactions. In a voice-based service, the system first needs to turn spoken words into text before it can interpret the request and produce an answer.
That initial transcription step is particularly important in financial services. A small error can change the meaning of a question, leading to an irrelevant or misleading response.
NatWest highlighted an example from its testing in which a Scottish colleague’s words “last quarter” were transcribed as “last water." Although the phrases sound similar to a speech-recognition system, they mean very different things in a banking context.
The challenge extends beyond regional accents. Merchant names, dates, financial terminology,, and informal descriptions of spending can all be difficult for speech recognition to handle consistently.
For a conversational banking service, understanding the intended question is not simply a matter of convenience. It helps determine whether the system retrieves the right information and presents a useful explanation of a customer’s finances.
How NatWest and NVIDIA tested the technology
The project began with testing involving 90 NatWest colleagues from different parts of the UK. Participants represented a broad range of accents, including Glaswegian, Northern Irish, Welsh, Yorkshire, Mancunian, Midlands,, and Southern English speech, alongside non-British accents.
Together, the participants recorded around 3,000 realistic banking questions. Rather than relying exclusively on scripted sentences, the researchers also asked colleagues to phrase questions naturally, reflecting the way people actually talk about money.
This distinction matters because people rarely speak to a banking assistant in the same way they would read a prepared sentence. They may shorten phrases, use regional expressions, change their wording midway through a question,, or refer to financial products and merchants by familiar names.
The researchers used the recordings to identify where speech-recognition models performed well and where they struggled with particular pronunciations, terminology,, or ways of phrasing a request.
NatWest then expanded the work using anonymized contact-center transcripts. These provided a wider range of real-world banking language and helped the teams improve recognition of merchant names, financial terms,, and everyday descriptions of spending.
The bank says the resulting system misheard 20% fewer words across the accents tested. This is a reported improvement from the project’s evaluation, rather than evidence that the system is error-free or that the same reduction will occur for every customer in every situation.
How NVIDIA Nemotron works with NatWest’s own AI
NatWest selected NVIDIA Nemotron automatic speech recognition technology for the speech-to-text component of Spending Insights.
Automatic speech recognition, or ASR, converts spoken language into written text. In a conversational banking system, that text is then passed to other AI components that interpret the request and determine an appropriate response.
NatWest uses its proprietary small language models to interpret the recognized text and generate relevant answers. This means the service combines speech recognition from NVIDIA’s technology with the bank’s own AI capabilities.
The distinction between these components is important. Speech recognition determines what the system believes the customer said. The language model then uses that transcription to understand the request and formulate a response.
Improving the first stage can reduce errors that might otherwise carry through the rest of the conversation. However, more accurate transcription alone cannot guarantee that the system will always interpret a request correctly or produce an appropriate answer.
NatWest and NVIDIA say they will continue optimizing the technology as conversational banking develops.
Spending Insights turns transactions into conversations.
The speech-recognition improvements support Spending Insights, a service designed to help customers explore their financial activity through natural conversations.
Customers can ask questions about spending, investigate trends, and receive personalized insights. The experience also supports follow-up questions, interruptions, and additional context, allowing people to refine what they are asking without starting a new interaction each time.
Responses combine conversational explanations with interactive charts and visualizations. The aim is to make transaction data easier to explore and understand, rather than requiring customers to manually examine individual payments or navigate multiple screens.
For example, a customer could ask about spending over a particular period and then follow up with a question about a specific category. The usefulness of that interaction depends on the system correctly identifying the period, category, and intent expressed in the original request.
NatWest previously introduced Spending Insights as part of a broader program of AI tools intended to help customers manage money and access support. The latest announcement focuses on improving the speech-recognition technology that underpins the experience.
The bank says the service is being rolled out through Royal Bank of Scotland, with access extended to 75% of its customers during the current phase. The rollout is ongoing, and wider group availability is planned.
Customer data remains within NatWest’s environment.
The collaboration also highlights how banks can work with external AI technology providers while maintaining controls over sensitive information.
NatWest says the contact-center transcripts used in the project were anonymized and remained within its secure environment. According to the bank, no NatWest customer data was shared with NVIDIA, and adaptations made using NatWest data are ringfenced to the bank.
These are important safeguards in a sector where financial information and customer conversations require careful handling.
However, anonymization and restricted data sharing should not be interpreted as eliminating every privacy or security risk. The effectiveness of such safeguards depends on how data is prepared, access is controlled, systems are monitored, and information is managed throughout the development and deployment process.
NatWest says its approach to AI is governed by its AI and Data Ethics Code of Conduct, which sets out principles for responsible innovation and protecting customers and their data.
For financial institutions, the broader lesson is that deploying conversational AI requires more than improving model performance. Banks must also establish appropriate data governance, security controls, and oversight as services become more sophisticated.
What the development means for conversational banking
NatWest’s work illustrates a growing focus in financial services. AI: making systems reliable in real customer interactions, rather than evaluating them only on general language tasks or controlled demonstrations.
Regional speech patterns and everyday expressions are part of how customers communicate. If an AI system performs well only when users speak slowly, use standardized language or follow scripted prompts, its usefulness in daily banking may be limited.
Improving speech recognition can help reduce friction for customers who prefer speaking to typing. It may also make it easier to ask follow-up questions and explore financial information conversationally.
Nevertheless, accuracy should be measured across a range of users and real-world conditions. Background noise, unfamiliar merchant names, ambiguous dates, and unexpected phrasing can still affect recognition. A transcription improvement is therefore one component of a larger effort to build trustworthy conversational services.
Banks will also need to ensure customers can verify important financial information, correct misunderstandings, and move to other support channels when a conversational system cannot answer confidently.
The next phase of AI in retail banking
The NatWest-NVIDIA collaboration reflects a broader shift from adding AI features to banking applications toward building more natural ways for customers to interact with financial services.
The potential value is not limited to voice commands. A conversational interface can help customers explore spending patterns, ask questions in stages, and receive information presented in a format that is easier to understand.
For banks, however, these experiences must balance convenience with reliability, privacy and appropriate customer support. A fluent conversation is not enough if the system misunderstands a question or presents an inaccurate interpretation of a transaction.
NatWest’s approach combines an external speech-recognition model with its own language models and banking-specific research. The use of recordings from colleagues with varied accents, followed by anonymised contact-centre transcripts, shows how domain-specific testing can help address limitations that general-purpose speech systems may encounter.
The reported 20% reduction in misheard words is a promising technical result, but the longer-term measure of success will be how consistently the technology performs for customers as access expands.


Comments
No comments yet. Be the first to share your thoughts.
Leave a comment