Speech Recognition Quality on Barge-In

The company I work for is creating an IVR application that utilizes speech recognition. One of the prerequisites is that during the asked question the user can barge in with the answer.

The issue we are coming up against is that the confidence interval for the speech recorded during barge-in is often very poor.

Extraneous data that may be useful:

[]We are using Lumenvox for our speech recognition[/]
[]We are running Asterisk[/]
[]When testing using an internal SIP connection the confidence interval is much higher than when we test using an outbound line.[/]
[]This occurs regardless of the complexity of the grammar files.[/]
[]Hardware echo cancellation is enabled and functioning.[/]

Anybody have any idea why this is happening, or what the fix could be?

Echo cancellers take time to train and will never be perfect.

There may be distortion products in the echo.

There may be anti-vox operation.

I’d certainly be surprised if voice recognition worked ell in non-ideal conditions.

Thanks for your reply David.

I’m not certain the problem lies strictly with echo cancellation, and was wondering if anyone else has had experience with speech recognition quality during barge-in.

I’ve tried it both with hardware/software echo cancellation on and off, with no change in the quality of the recognition.

Any other thoughts?