Is voice recognition bad on cell phones?

[ban]

The voice recognition is very bad, especially on cell phones. Are there techniques to improve it? How do you limit the grammar (beyond type="number") when the allowable range is between 80 and 500 ?

[ban]

The voice recognition of numbers is not very good at all. In particular, I cannot get the system to recognize the number 80. (on a normal land line telephone.)

Has the system been trained at all? Is there more training/tuning that could be done? Are there more sophisticated configurations or speech recognition engines that could be used?

[support]

The voice recognition of numbers is not very good at all. In particular, I cannot get the system to recognize the number 80. (on a normal land line telephone.)

It's not so much a problem with speech recognition as it is an IVR issue with signal to noise ratio...If you are attempting to use speech recognition in a fairly noisy environment, this can impact 2 aspects of the IVR system:

1) The bargein and silence threshholds will cause the recording to start prematurely, or allow trailing noise into the IVR application.

2) The ambient decibel level will be high enough such that it becomes difficult to distinguish between speech in the foreground and noise in the background. This makes reliable speech detection almost impossible to achieve
on a consistent basis.

Has the system been trained at all? Is there more training/tuning that could be done?

The speech engine has been trained, but more training will provide better results. Plum will soon be seeking out volunteers to help record voice input from a variety of sources, so that the recognition will have a broad range of input samples to train on. This should also make for drastic improvements. Is this something you'd be interested in participating in?

Are there more sophisticated configurations or speech recognition engines that could be used?

Yes, there are other commercially available speech recognition engines that we have successfully integrated into our IVR platform. Keep in mind that our IVR hosting site uses an open source speech recognition platform, and that other options are available such as IBM Speechworks.

Hope this helps,

Plum Support

[ban]

Thank you for your input.

However, I am working in a near-silent office so the signal-to-noise ratio should not be an issue.

The system seems to recognize numbers like 100, 200 with no problem, but 80, 85, 65, it cannot recognize at all. Multiple people here have tested it, both females and males.

I am glad to hear that more training will be done, but still disappointed that recognition of basic numbers is so poor.

[support]

Generally speaking , you should not have problems with numbers in a quiet environment. If you post an IVR code snippet showing the vxml you use to perform your tests , we would like to try replicating your results.

sincerely,


Plum Support

[ban]

Here is the code. I was successful getting the system to recognize 85, but still no luck on 80. (Again, from a landline telephone, in a near-silent office.)

I have one other concern. The dial in number is set to a URL that contains the code shown below. The first time I called it today, I got a grammar error. Without changing anything, I called again. It worked fine. Any idea what the problem is?

Code: Select all

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE vxml PUBLIC "//The Plum Group//Voice Markup Language 2.0//EN" "http://vxml.plumgroup.com/vxml-221.dtd">
<vxml version="2.0" xmlns="http://www.w3.org/2001/vxml">
  <form id="id" scope="dialog">
    <block>
      <audio src="audio/welcome.wav">hello</audio>
    </block>

    <field name="field89" type="number">
      <prompt bargein="false">
      <audio src="audio/nonexistentfile.wav">
        What is your diastolic blood pressure?
      </audio>
      </prompt>
      <nomatch>
        <audio src="audio/sorry.wav">
          Sorry I didn't understand
        </audio>
        <audio src="audio/useKeyboard.wav">
          You can also enter the value using the keypad.
        </audio>
        <reprompt/>
      </nomatch> 
    </field>

    <field name="confirm89">
      <grammar type="application/x-jsgf">(yes|1) {yes} | (no | 2) {no}</grammar>
      <prompt bargein="false">
      <audio src="audio/nonexistentfile.wav">
        You entered <value expr="field89" />
      </audio>
      <audio src="audio/isCorrect.wav">
        Is this correct?
      </audio>
      </prompt>
      <filled>
        <if cond="confirm89=='no'">
          <clear namelist="field89 confirm89" />  
          <goto nextitem="field89" />
        </if>
      </filled>
    </field>

    <block>
    <submit next="http://companywebsiteurl/response.xml" 
      namelist="field89 " method="get" 
      enctype="application/x-www-form-urlencoded" />
    </block>
  </form>
</vxml>

[support]

Hello:

Most likely the reason why you are getting different results without changing the vxml you provided is due to the <submit> tag :

Code: Select all

 <submit next="http://companywebsiteurl/response.xml"
      namelist="field89 " method="get"
      enctype="application/x-www-form-urlencoded" /> 

The response.xml page is probably being modified, and when you attempt to submit information to it, you get an IVR error because the response.xml page is malformed. Hard to say for certain, but this is most likely causing it. Whenever the page works intermittently without changing the vxml document it almost always means that one of the documents you submit to has a problem in it.


P.S. If you are having another IVR issue, please post it as a separate thread in the forum for clarity and for the sake of other forum users


hope this helps,

Plum Support

[support]

Hi ban:

Try your speech recognition IVR code again; Plum just recently deployed a new IVR platform release on the IVR hosting environment (v2.4.2.7), and you may see an improvement in speech detection quality.

Check the news posts on the forums main page before you log in for more details.



kind regards,

Plum Support staff