NatWest has worked with Nvidia to improve speech recognition in its conversational AI services so they better understand accents and phrasing across the UK. The work supports the new Spending Insights tool, which lets customers ask about their spending by voice or text. Tests showed 20% fewer misheard words across the accents checked. The tool is already available to 75% of Royal Bank of Scotland customers.

What it means: Voice banking sounds simple, but in practice it often fails when the system does not understand local accents, idioms or the way people talk about money. A 20% cut in recognition errors can be the difference between a tool people use and one they abandon. The NatWest case shows that general speech models need tuning with local data to work in real conditions. For Greece the issue is even sharper. Greek is a small language for the large AI providers, with regional accents from Crete to Thessaloniki and many words mixed with English terms. Greek banks and companies planning voice assistants or call-centre automation should test accuracy with real customers from different regions and age groups, not just in the office. Customer recordings are also personal data, and using them to train models needs a clear legal basis under GDPR. Accessibility matters too, since voice helps older people and those with sight loss. A good starting point is a small pilot in one service line, with careful measurement of errors and customer satisfaction, before rolling voice features out more widely across all channels.

Read the original on NatWest Group

Related service: myDATA integrations

Related stories