Stories
Slash Boxes
Comments

SoylentNews is people

This Speech Recognition Code is 'Just As Good' as a Pro Transcriber

posted by martyb on Thursday October 20 2016, @04:45AM   Printer-friendly
from the nature'll-anguish-wreck-ignition dept.

Phoenix666 writes:

Microsoft on Tuesday said that its researchers have "made a major breakthrough in speech recognition."

In a paper [PDF] published a day earlier, Microsoft machine learning researchers describe how they developed an automated system that can recognize recorded speech as well as a professional transcriptionist.

Using the NIST 2000 dataset of recorded calls, Microsoft's software performed slightly (0.4 per cent) better than the error rate the company attributes to professional transcriptionists (5.9 per cent) for the Switchboard portion of the data, in which strangers discuss a specified topic.

There goes your bright future as a court recorder...

Original Submission

 
This discussion has been archived. No new comments can be posted.
This Speech Recognition Code is 'Just As Good' as a Pro Transcriber | Log In/Create an Account | Top | 31 comments | Search Discussion
Display Options Threshold/Breakthrough Mark All as Read Mark All as Unread
The Fine Print: The following comments are owned by whoever posted them. We are not responsible for them in any way.

  • (Score: 2) by RamiK on Thursday October 20 2016, @11:57AM

    by RamiK (1813) on Thursday October 20 2016, @11:57AM (#416588)

    Might be real. Past attempts had humans trying to analyze speech and produce algorithms. This time, after years of failure, Microsoft dumped the data-set on a neural net and let run until it worked.

    There been a few discoveries in medicine following this approach already so it's not all that implausible.

    --
    compiling...
    Starting Score:    1  point
    Karma-Bonus Modifier   +1  
    Total Score:   2  

  • (Score: 0) by Anonymous Coward on Thursday October 20 2016, @08:19PM

    by Anonymous Coward on Thursday October 20 2016, @08:19PM (#416916)

    Heh, yeah.
    I remember one spectacular failure with Visduh's speech recognition as demo'd by Gates.
    Dear aunt, let's set so double the killer delete select all [google.com]

    Microsoft dumped the data-set on a neural net and let run until it worked

    I was wondering if they're use multiple methods, with each running on its own core, then taking a majority vote to see if multiple methods arrived at the same answer.

    -- OriginalOwner_ [soylentnews.org]

    • (Score: 2) by RamiK on Friday October 21 2016, @11:35AM

      by RamiK (1813) on Friday October 21 2016, @11:35AM (#417192)

      I was wondering if they're use multiple methods

      Unnecessary. They could have fed the net the original algorithm and the data-set and have it approximate an improved solution.

      Human speech is already a neural phenomenon. Compared to using a neural net which is designed to imitate the brain, using the approximation methods you'd use for differential equations is inherently inefficient at producing \ correcting the algorithm. Now that they have parity with humans they might look into silicone optimizations and the likes... But started off with a neural net should have been the obvious thing to do.

      I'm guessing it also explains why neural nets are so good for image recognition if we're using human capacity as our test unit.

      --
      compiling...