News

Oxford Develops AI System for Significantly Improved Lip-reading

Using thousands of hours of BBC News programmes including Breakfast, Newsnight, Question Time and more, scientists at Oxford have developed an artificial intelligence system that can lip-read better than humans. The project has been developed in collaboration with Google’s DeepMind AI division, dubbed the “Watch, Attend and Spell” system, now boasts a 50% lip-reading hit-rate.  In comparison, the same lip-reading tests shown to professional lip-readers only have an accuracy of 12%.

The neural network using speech recognition and image recognition algorithms is able to gather 17,500 words for its vocabulary from examining 118,000 sentences in the clips. Since it is fed mostly news programmes, it has grown to associate the likelyhood of certain words following another in a particular topic such as “minister” after “prime”, but this also means that it is limited and cannot recognize many words that are not spoken by newsreaders.

As good as the system is now, the researchers still expect a lot of work needed before it can be put to practical use. Many groups, such as those who are advocates of the hearing impaired, are very excited about the development however.

“AI lip-reading technology would be able to enhance the accuracy and speed of speech to text,” says Jesal Vishnuram, Action on Hearing Loss technology research manager. “This would help people with subtitles on TV, and with hearing in noisy surroundings.”

The next objective for the Oxford researchers, is to make the system work in real time. For now, it is only able to operate on full sentences from recorded video. According to Joon Son Chung, a doctoral student at Oxford University’s Department of Engineering, this is actually a simpler task than refining the accuracy of the AI system, so it is not as challenging as it sounds.

Ron Perillo

Disqus Comments Loading...

Recent Posts

Nvidia’s GeForce RTX 5090 Possible Price Revealed

According to a new report, the GeForce RTX 5090 GPU will be very expensive. It…

9 hours ago

AMD Krackan Processor with 6 Zen 5 and Zen 5c Cores for Budget AI Laptops Leaked

A new AMD processor in the form of an engineering model has been leaked in…

9 hours ago

SK Hynix Begins Production of First 321-Layer NAND Chips

SK Hynix has claimed to be the first company to mass-produce 321-layer NAND memory chips.…

9 hours ago

Trust Gaming GXT 609 Zoxa 2.0 PC Speakers

SOUNDS GREAT – Full stereo sound (12W peak power) gives your setup a booming audio…

13 hours ago

PowerA Wired Controller for Nintendo Switch

Special Edition Yoshi design Ergonomic controller shape with Nintendo Switch button layout Detachable 10ft (3m)…

13 hours ago

Logitech G Saitek PRO Flight Rudder Pedals

Fluid Motion: These flight rudder pedals are smooth and accurate that enable precise control over…

13 hours ago