noise problem for ivr prompt

[sheeja.v]

Hi,

We have set sensitivity to 0.0 and confidence level to 0.8 but still the plum voice takes outside noise as input.

For these words 'two' , 'confirm', 'ok' any outside noise is being taken as match.
For the words 'scratchpad' and 'notepad' even saying pad is matched with notepad.

Please see below sample vxml

<?xml version="1.0"?>
<vxml version="2.0">
<property name="sensitivity" value="0.0"/>
<property name="confidencelevel" value="0.8"/>
<form>
<field name="username" type="digits">
<prompt>
please enter six digit username
</prompt>
<filled>
you enter <value expr="username"/>
</filled>
</field>
<field name="usernameconfirm">
<grammar>
confirm {1} | re-enter {2}
</grammar>
<prompt>
To confirm say confirm.
To re-enter say re-enter.
</prompt>
<filled>
you selected <value expr="usernameconfirm"/>
</filled>
</field>
</form>
</vxml>

In the place of confirm we have tested with 'ok' but the results are same.

[support]

Hi,

There are a few things we would like to mention about this issue.

Firstly, we would recommend keeping the sensitivity threshold around 0.3 because at any point lower you're desensitizing the ASR enough that you might be missing some of the incoming speech.

Additionally, to get an idea of how accurate your responses are you can use the confidence application variable to log or playback what the confidence level was for your ASR match. Details about this variable can be found here:

http://www.plumvoice.com/docs/dev/voice ... confidence

We tested your script by coughing and shuffling around some papers but these were rejected with a nomatch as expected. We also tried saying portions of your grammar phrases such as "re" and "firm" but were also rejected the majority of the time. Note that the platform is trying to do a closet match to your defined grammar, so it's not totally uncommon for a partially spoken phrase to match one of your specified grammar keywords (with varying levels of confidence). Again, you can use the confidence application variable to get an idea of how restrictive you would like to set your confidence level to only accept speech that you think is acceptable.

One additional thing to note is that, unless you necessarily need to use ASR, you can also restrict your input to DTMF only. You can do this by setting this property within your script:

Code: Select all

<property name="inputmodes" value="dtmf"/>

You can alternatively specify this property globally within your entire application using a root document. Anything you declare in your root document will apply to any VXML file that links to your root document.

An example of a root.vxml document:

Code: Select all

<?xml version="1.0"?>
<vxml version="2.0">
    <property name="inputmodes" value="dtmf"/>
    <property name="sensitivity" value="0.3"/>
</vxml>

You can then include this in any vxml script as such and that script will have all of those same properties applied:

Code: Select all

<vxml version="2.0" application="root.vxml">

Please let us know if you have any additional questions.

Regards,
Plum Support

[aditya.ramam]

Hi, Thanks for the feedback. We are working on testing a "voice-input" driven application where ASR is the key and having dtmf only as input is not an option. Our tests, mainly, when there are background noises such as door opening etc combined with kind of "sounds" you have mentioned have repeatedly failed. Please advise

[support]

Hi,

As we had previously mentioned, extraneous sounds can interfere with ASR. Background noises such as opening doors, like you mentioned, if loud enough, will be picked up and cause a nomatch. This is expected behavior and we can only suggest you test in environments were background noises are minimal.

Regards,
Plum Support