NatWest and Nvidia tune speech recognition to UK accents and cut misheard words by 20%

The adapted system got 20% fewer words wrong and posted a lower overall Word Error Rate, the announcement said ^[14]. Take that as a relative cut and the remainder is easy to size: if the old system misheard five words, the adapted one mishears four, so four of every five original errors remain ^[15].

The remainder matters most where the testing found the most trouble. Merchant names, dates and financial terminology proved particularly difficult, despite being central to interpreting a customer's query, and minor transcription errors can change what a request means ^[5]. A word-level rate counts a mangled merchant name as one error among all the others. A spending query answered for the wrong merchant or the wrong month would arrive with the same charts as a correct one, since replies combine conversation with interactive charts ^[11].

The panel behind the accent work was 90 NatWest employees ^[3] who recorded around 3,000 banking questions between them ^[4], about 33 each ^[12], both reading scripted versions aloud and rephrasing them in their own words ^[4]. That is a tidier sample than the customer base. The announcement does not say whether the 20% was measured on that panel, on the later contact-centre transcripts ^[6] or on live traffic.

The allocation is the more interesting part. NatWest uses Nvidia's Nemotron ASR, from an open-source model family, for the first stage and its own Small Language Models for the second ^[2]. The transcripts were anonymised and stayed inside the bank, no customer data went to Nvidia, and any adaptations made with NatWest data are ringfenced to the bank ^[7]. The bank is therefore paying for panel testing, transcript preparation and its own models, and it is not sending its data to a vendor. Nvidia's gain, in the account of Aser Blanco, its Global Head of Banking, was applying Nemotron models to real-world banking use cases and identifying where they could be improved ^[8].

We think the 75% access figure is firmer than the 20% accuracy figure. Access is a rollout fact: 75% of Royal Bank of Scotland customers can use Spending Insights now ^[1], 25% cannot yet ^[13], and the wider NatWest Group follows ^[1]. Accuracy is a claim about an undisclosed base. Two outcomes would make us wrong in opposite directions. A before-and-after on merchant names and dates showing the gain holds where errors cost most would mean we underrated the work, and live traffic showing a gain well under 20% would mean we overrated it.

According to the announcement, Maja Pantic, NatWest's Chief AI Research Officer, said customers increasingly use AI in daily life, and that for those who opt to talk to it, the technology has to cope with the range of accents and ordinary speech found around the UK ^[9]. The two companies plan to keep refining the recognition as conversational banking services develop ^[10].