Microsoft's Machine Transcription Achieves Parity With Humans


Microsoft's experimental speech recognition software achieved the lowest-ever recorded error rate for machine transcription this week, matching the abilities of a human translator.

In a research paper published Monday, Microsoft engineers reported that their software had achieved a word error rate (WER) of 5.9 percent, similar to the results of people who were asked to transcribe the same conversation.

As recently as five years ago, even the best speech recognition systems still had word error rates of 20 to 25 percent. Microsoft boasted of a 6.3 percent WER last month, but today's report is the first that recorded a rate below 6 percent.

In a blog post, the company stressed that the achievement does not mean its software was able to transcribe speech perfectly, since even humans can't do that. Nevertheless, it is a win for neural network research: the neural language model used in the software learns the relationship between words, not just their sounds. That means the language processing engine can recognize synonyms—if a speaker uses the word "fast," it will look for "quick" and other similar words to follow.

Microsoft did not announce when the software used in the study will make its way into commercial products.

Google, Apple, and other companies have recently been touting their own efforts to teach neural networks how to recognize speech and other sound patterns. In May, Google showed off its Magenta project, which applies the neural network idea to composing music.

Meanwhile, researchers have also tested machine speech recognition against a far more modern means of communication: texting. Researchers at Stanford University devised an experiment in August that pitted Chinese tech giant Baidu's speech recognition software against 32 texters, ages 19 to 32, working with the built-in keyboard on an Apple iPhone. Baidu's software was not only three times faster than the human typists, it was also more accurate.

0 Comment

Leave a Reply

Captcha image


  • 5300c769af79e

    Report: T-Mobile Uncarrier 11 is "Stock Up," Customers Can Earn $TMUS Stock

    According to a source of VentureBeat, T-Mobile’s Uncarrier 11 event is geared towards getting postpaid customers invested more in the network they use.As reported, Uncarrier 11 will introduce Stock Up, a program that offers actual T-Mobile stock to customers.
  • 5300c769af79e

    Angel Diaz - Authors & Columnists

    5/31/2016Dr.Angel Diaz of IBM co-chaired and co-authored the first XML standard in 1998.
  • 5300c769af79e

    Twitch Launches PS Vita App

    Like the PlayStation 3 app, Twitch on Vita allows you to watch live and recorded video of top games, players, and e-sports events; follow channels and games; and discover new PlayStation broadcasters and other streamers across the platform.Just grab the portable console, download the app from the PlayStation Store, and turn on the data connection.
  • 5300c769af79e

    Jazz Up Photos With Twitter's Searchable #Stickers

    The new photo-editing tool lets users add digital accessories, emoji, and props to photos uploaded to the social network.When editing a photos, tap the smiley-face icon at the bottom of the screen to access a library of labels; skim through various categories—accessories, animals, food, technology, flags, etc.