Recognizing Speech With AI

October 2020

Researchers at the Karlsruhe Institute of Technology (KIT) have developed an artificial-intelligence-based system they say recognizes human speech better than human beings do.

“When people talk to each other, there are stops, stutters, hesitations, ‘er’s’ and ‘hmm’s,’ laughs and coughs,” Alex Waibel, Professor for Informatics at KIT told the Internet platform ArXiv.org. “Often, words are pronounced unclearly.”

Waibel says that his institute’s new system reduces the error rate to 5.0 percent – compared with 5.5 percent for humans. The latency of the system, he added, has been lowered to one second, a new record for quality voice recognition tools.

The system was developed from the so-called Lecture Translator in use at the university since 2012. It’s employed to directly translate lectures for foreign students.

© KIT