In partnership with

Speech recognition: a solved problem, right? Not according to today’s startup, which looks to fill the gaps the established players can’t or won’t address, with a clever and highly efficient new approach.

Read on to discover FlexSR.

– Martin

Your first HR system, implemented right

Rolling out your first HR tool? Get a step-by-step guide to avoid common mistakes, drive adoption, and build a scalable HR foundation.

Download the guide

FlexSR wants to take on the giants of speech recognition by thinking really small

Speech recognition tech is an increasingly important part of everyday life.

Whether we’re talking to assistants inside our phones, being screened for customer service phone calls, or having our work meetings transcribed by AI assistants, software is turning our voices into data all over the place.

But that speech recognition isn’t anywhere near perfect. FlexSR is a startup taking a fresh approach to solve this problem.

“There are constraints and restrictions with current technologies from amazingly big companies that are hugely resourced with great processing power and data stores. They still get things wrong when there are strong dialects, when people are switching languages, or when someone has just invented a whole bunch of words and slang and phrases,” says Stephen Fulton, CEO and co-founder of FlexSR.

The startup is a spinout from the University of Oxford, based on technology that allows the creation of a small, language-independent speech recognition model that can support use cases that other approaches struggle with. What’s more, they say it works in any language without training.

“It approaches speech recognition from a more human, linguistic way. As humans, we make these little sounds that combine to make up the bigger sounds that make up words, which is very different from trying to create an incredibly huge model that tries to build and model lots and lots of different conversations to match them to a best fit,” Fulton says.

“No matter what language you speak, all humans make these sounds, these phonological features. These combinations make the phonemes that make words, and there are only between 19 and 21 of these sound features that are required to form any of the sounds that make any words in any language…

“So the job of identifying these small sound features in an audio signal is faster and much lighter than [other approaches].

“The principles of these phonological features and how they combine to make up phonemes, has evolved over the decades, but it wasn't really put into a cohesive system that you could apply in the real world to these speech recognition problems.”

Another benefit of a small model is that it can be run efficiently on-device without any need for an internet connection. Fulton says it could run a smartwatch, for example.

The story so far

Fulton has worked in tech throughout much of his career. In the 90s, he helped scale a UK startup, Eyretel, to an IPO as sales director.

He’s also worked in and around the speech recognition field a lot. He spotted the need for technology to monitor phone calls in the financial sector for regulatory purposes at the beginning of the century.

This was at a time when online banking apps were yet to become commonplace, and trades and transactions typically involved voice-based conversations.

Throughout his career, he says he has kept up-to-date with the latest speech recognition technologies, but has found them lacking.

“I was getting frustrated that the art of the possible just wasn’t good enough, and I couldn’t believe that there wasn’t a better way of recognising the words that people were saying from different languages,” Fulton says.

And so he began to look at university research and discovered something at the University of Oxford that could fit the bill. He came across the work of Professor Aditi Lahiri CBE and Professor Henning Reetz and offered to commercialise their research.

FlexSR was spun out in February 2024. It has now reached the point where, Fulton says, a UK police force has agreed to take part in a production pilot project. This will allow FlexSR to reach MVP stage with a product for the force to deploy that will also be usable by other customers.

Fulton says he has received interest from the investment banking and defence markets, which the tech will be ready to address once it is built out for the police.

“This is something that can scale globally… We’re able to package it up in a small software container… then we’ve got a really open door, not just to the end users, but to the market, because we can then work with market participants that have built solutions,” Fulton says.

“They’ve built something that will take speech in, that will do some type of processing, and has an output, but they’ve got a missing component in terms of the capability and accuracy that we can give them with FlexSR in a container. So that way, we’re opening up opportunities with partners. We’re hitting gaps instead of having a head-on fight with Google, Amazon, and Microsoft.”

Subscribe to Premium to read the rest.