Google’s DeepMind AI has been trained to lip read more accurately than a professional human lip reader after watching hours of video footage.
One day we’ll all be able to thank Google for plugging in the overlord – researchers from Google’s AI division (DeepMind) and the University of Oxford have trained artificial intelligence to become the most accurate lip-reading software yet made.
While a professional human lip-reader has an accuracy rating of 12.4% when annotating video footage, Google’s DeepMind AI has achieved an accuracy rating of 46.8%.
Read: Google CEO Sundar Pichai outlines his vision for artificial intelligence
Comparatively, researchers at the University of Oxford used related techniques to create a lip-reading program called LipNet that achieved 93.4% accuracy versus 52.3% professional human accuracy just one month ago. You can read more about that study and the metrics used to determine accuracy here.
However, LipNet was tested on footage created especially for research purposes which used volunteers speaking formulaic sentences. DeepMind has been given a loftier challenge – annotating unscripted conversations from politics shows that have aired on the BBC.
DeepMind has thus far ingested over 5,000 hours of video footage from TV shows including Newsnight, Question Time, and World Today – the videos include 118,000 different sentences and 17,500 unique words.
DeepMind researchers have cited that the program could help hearing-impaired individuals understand conversations, or could be used to annotate silent films. Similarly, the AI’s lip-reading prowess could help to improve digital assistants such as Siri, Cortana, or Google Assistant were they able to leverage a smartphone camera to analyze speech.
Read: The smartphone race is over, but the AI war has just begun
What are your thoughts? How else could we use an AI trained to lip-read for beneficial purposes? Be sure to let us know your thoughts in the comments below!
Follow Bryan Smith on Twitter: @bryansmithSA
Source: The Verge
