Microsoft Computer Can Transcribe Conversations As Well As a Human

SkyNet may be only a few steps away.

Microsoft researchers say they have come up with a technology that can recognize words in a recorded conversation just as well as an actual human being would, claiming victory over a problem computer scientists have been working on for decades.

The company’s Artificial Intelligence and Research division announced Tuesday that its speech recognition system had reached a word error rate of 5.9 percent, about equal to that of professional transcriptionists.

Researchers used a series of recorded phone conversations known as the Switchboard test. Developed in the 1990s by the National Institute of Standards and Technology, the test has long been used by tech companies to check the accuracy of speech recognition software, according to Business Insider. NIST verified the results of Microsoft’s test.

Microsoft’s speech and dialog research manager, Geoffrey Zweig, emphasized that while the computer can accurately recognize words, it doesn’t necessarily know what they mean.

“The next frontier is to move from recognition to understanding,” Zweig said in a press release.

The team used neural network models to teach the computer to recognize sets of similar words. The deep learning system they developed, dubbed the Computational Network Toolkit, is available on GitHub.

The breakthrough comes at a time when Microsoft has been focusing intently on AI. The company last month announced a major reorganization that consolidated a number of pre-existing teams in a new Artificial Intelligence and Research division.