locked
Recognizing from wave data RRS feed

  • Question

  • Hello,

    I want to evaluate the recognition performance of the speech server
    to compare against other engines.
    To do so, I want to recognize a set of wave files.

    Is there a "setInputToWave" kind of function for MSS?
    The set is too large to do it with the sipphone under VS, so
    any ideas to make the test automatic are welcome!

    Thanks in advance.

    -Ronaldo
    Wednesday, March 11, 2009 12:43 PM

All replies

  • You can do this by building two small applications; one for outbound and another for inbound. The outbound application dials the inbound application and plays an audio file. By logging both the file played from the outbound application and the recognition string from inbound application you can determine how accurate the recognition result was.

    One thing to keep in mind when testing accuracy of any recognizer is the size of the grammar you’re testing against. The larger the grammar the more difficult it is to recognize against. A grammar with 100 entries is going to result in more accurate recognition than one with 10,000 records. You need to make sure you’re modeling both sides; the audio your application will likely receive from callers and the grammar it will likely have in production.


    - Marc LaFleur
    Wednesday, March 11, 2009 1:21 PM
  • Thanks Marc!

    Can I do the calling and answering on the same xp server
    with the evaluation version?

    This is my first try with MSS and this kind of application development, so
    I am a bit uneasy with the development part under VS and .net
    and I am on a tight schedule...

    I'll check the sample applications to see if I can implement your solution.

    I really appreciate your help.

    -Ronaldo

    Thursday, March 12, 2009 6:28 PM
  • You can use the Speech Server Developer Edition. A common misunderstanding (and an understandable one) is that the Speech Server Developer Edition is the full version of Speech Server. To license it for production use you’ll need an OCS 2007license but for non-production (development/test) you do not need one.

    As for calling itself, I’m not sure. There isn’t any reason I know of that it couldn’t but I’ve never tried.  

    I’ve got a blog entry covering some common outbound issues at http://shrinkster.com/154j that you may want to read before starting.  I’ve also posted on that addresses an issue calling from one Speech Server to another involving 302 Redirects at http://shrinkster.com/154k.  

    Hope this helps some.


    - Marc LaFleur
    Friday, March 13, 2009 3:23 AM