FlexSR is a University of Oxford spinout building on-device speech recognition. Our worldwide patented engine matches phonological features instead of predicting words, so it runs on hardware as small as a Raspberry Pi with no cloud connection and no retraining per language or accent. We work above existing ASR as a verification and spotting layer, confirming what was actually said.
Who we are
Our Vision
To create a world where every spoken word that matters is captured, confirmed, and turned into action at the moment it is said.
Our Mission
Verifying words. Nothing should be misheard.
The Board
Stephen Fulton
Lead Researcher
Professor Lahiri has spent over 30 years in linguistics in the field of speech recognition. Her career has been recognised and in vast amount worldwide major honours and is known as a legend in her field. Currently, she is a Fellow of the British Academy, an Honorary Life Member of the Linguistic Society of America, and a Fellow of Somerville College.
Prof. Aditi Lahriri
Senior Researcher
Professor Henning is the co-inventor of FlexSR. With decades of experience, he was the Professor of Phonetics at the Institute of Phonetics, Goethe University Frankfurt, Germany.
Prof. Henning
Product Director
John has over 25 years of experience working with ASR and related technologies in various technical roles. He specialises in digital signal processing and speech technologies. Conducted R&D on speech technologies including speech enhancement, biometrics and ASR.
John Jacob
Enterprise Director
Peter is a founder in successful high-performance enterprise data processing companies.
Built DealHub, which was acquired by Markit in 2015. (Nasdaq: MRKT), a leading global provider of financial information services.
Peter Kriskinans
Backed by
The Executive Team
Professor Aditi Lahiri
Lead Researcher
Professor Henning Reetz
Senior Researcher
Stephen Fulton
Co-founder and CEO
Peter Kriskinans
Enterprise Director
Product Director
John Jacob
Linguistic and Coding Analyst
Dr. Emily Lindsay-Smith
Our Story
Aditi Lahiri spent years at the Max Planck Institute in Nijmegen, then at Konstanz, working on how the brain turns sound into words. Her question was a simple one. Nobody says a word the same way twice, not even their own name, so how does a listener understand it every time?
Before 2008
Aditi came to Oxford as Professor of Linguistics and founded the Faculty of Linguistics, Philology and Phonetics.She set up the Language and Brain Laboratory on Wellington Square, and Henning Reetz kept working with her on the same question.
2008
2011
Her first ERC Advanced Grant funded five years of work on what the brain actually stores when it stores a word.The answer was not whole sounds. It was phonological features, the small properties that separate one sound from another. Voicing, whether your vocal cords are buzzing, is the whole difference between a b and a p.
2015
An ERC Proof of Concept award paid to turn the theory into software, and Aditi’s team built FlexSR from scratch in six months.It was trained to recognise a universal set of nineteen phonological features and combine them into speech sounds. That is why it never needed thousands of hours of recordings for each new language.
2016
Henning Reetz, Vipul Arora and Aditi published the feature detection work at an IEEE machine learning workshop, and the mispronunciation work at Interspeech the year after.Patents were filed. At this point FlexSR was still a research tool, running English and German, sitting in a university lab.
Through the same years Stephen Fulton was selling speech recognition to compliance and investigation teams in the financial markets. Their machines listened to thousands of calls a day looking for misconduct.The big systems coped with ordinary conversation and then lost the words that mattered. A dialect nobody had trained on. Slang invented last month. A speaker switching language halfway through a sentence.It seemed unlikely that nobody had found a better way.
February 2024
Stephen found the Oxford work and realised the professors had already solved the problem he had been describing for years.FlexSR Limited was incorporated in February 2024, with the University and Oxford University Innovation as shareholders, and Aditi and Henning as inventor founders.
2025
The research code was rebuilt for production, paid for with angel investment and grants of around $320,000. The Defence and Security Accelerator assessed the technology and rated it Fundable.
2026
Police Scotland began a live pilot, in a service taking around three million 999 and 101 calls a year, against a dialect that defeats most transcription outright.The Ministry of Defence wrote letters of support. A British Army lexicon went under test on a Raspberry Pi, a board the size of a credit card, with nothing leaving the device.
June 2026
The ERC awarded another Proof of Concept grant, this one for speakers who switch between languages mid sentence.That is the exact failure Stephen had been describing in the markets years earlier. It is now a funded Oxford research project.
2027
The plan is a general release in the first half of the year and first revenue in the second, on a seed round of around $2 million to take the pilots to paying customers.