Microsoft's System Has a Word Error Rate Equal to That of Humans

By: Joey Haar - Oct 20, 2016
References: blogs.microsoft & theverge
A team of engineers and researchers from Microsoft Artificial Intelligence and Research have designed a speech recognition system with a word error rate equitable to that of humans.

With more and more technologies and systems in place that listen to users' voices for commands, including Microsoft's own Cortana, speech recognition systems need to be incredibly accurate. The new system from Microsoft is so accurate that it has achieved parity with professional human transcriptionists. The system's word error rate, or the percentage of incorrect words in a transcription, is 5.9 percent, which is equal to the average for humans.

The researchers attribute the achievement to their use of neural networks, since these programs group words by meaning instead of sound. Thus, the word "fast" in the computer's brain is similar to "quick" but not to "feast,' for example.