With Voice Activity Detector (VAD), Avaya Experience Portal can detect the presence or absence of human speech. VAD can help detect if the sound is background noise or actual speech (speech energy).
VADs are used to determine when to send the audio stream to cloud speech resources such as Dialogflow, Google Speech, or Nuance Dialog as a Service (NuanceDaaS). When AEP VAD detects what it determines as speech energy, AEP then sends the audio stream to the cloud speech resource.
A properly tuned VAD reduces the incidence of falsely interrupted announcements when barge-in is enabled.
The AEP VAD configuration can enable and disable the local VAD and use the cloud speech resource instead. The AEP VAD configuration supports the following features:
Disable AEP local VAD - Disabling the local VAD and immediately streaming all audio to the cloud speech resource (Dialogflow, Google Speech, or NuanceDaaS), thereby using the VAD of the cloud.
Hybrid option - The local VAD is used to detect the start of speech, but the cloud speech resource is used for the rest of the utterance.
Important:
Disabling AEP local VAD or using the Hybrid option leads to increased network usage as the audio is streamed immediately to the cloud speech resource. It also leads to increased charges from the cloud speech resource provider since the cloud speech resource VAD is used.
AEP implements the following VAD algorithms:
GMM - Gaussian Mixture Model
Inqa - Speech detection using Call Progress Engine
MeanSquare - Original VAD algorithm (deprecated VAD)
VAD Configuration
VAD configuration consists of VAD selection and VAD tuning. VAD configuration can be applied on a global system-wide basis or a per-application basis.
AEP supports separate VAD configurations for each application. This overrides the global system-wide VAD configuration for that application.
When configuring an application for the cloud speech resource provider (Dialogflow, Google Speech, or NuanceDaaS ASR), you can configure the VAD parameters for that application.
VAD Mode
VAD Mode determines whether local VAD, the VAD of the cloud speech provider, or a combination of both VADs are used.
The VAD mode values are the following:
<None> - AEP local VAD is enabled and uses the default GMM VAD or the system-wide VAD type specified in mppconfig.xml.
Disabled - AEP local VAD is disabled and the audio stream is sent immediately to the cloud provider.
Enabled - AEP local VAD is enabled and uses the selected VAD type algorithm to detect speech energy. Once speech energy is detected, AEP sends the audio stream to the cloud provider.
Hybrid - AEP local VAD is enabled for the initial response in every call but is disabled for all other utterances for the rest of the call. Therefore, the local VAD is used first, and then the cloud provider VAD is used for the rest of the call.
VAD type
VAD type determines which local VAD algorithm is used.
The VAD type values are the following:
<None> - Default GMM VAD or the system-wide VAD type specified in mppconfig.xml is used.
GMM - Gausian Mixture Model VAD algorithm is used.
Inqa - Inqa Call Progress Engine VAD algorithm is used.
For more information on Avaya Experience Portal Voice Activity Detector (VAD), see Avaya Experience Portal Dialogflow White Paper.
