Skip to main content

Best Practices for Build a VoiceX AI Agent

Managing Latency in VoiceX Conversations​

The average voice-to-voice latency (time between the user’s last word and the bot’s first word) in a VoiceX conversation is approximately 1.5 seconds. Below is the latency breakdown:

  • Final Silence Duration: 200 ms
  • STT (Speech-to-Text) Transcription: 0 - 300 ms
  • Platform Latency / LLM First Sentence: 1 - 1.5 s
  • TTS (Text-to-Speech) Time to First Chunk: 0 - 1 s
  • Total Latency: 1.5 - 2.7 s

If latency exceeds 1.5 - 2.7 seconds for multiple turns, raise a DevRev ticket.

Enabling Interruptions​

Enable Interruptions in Dynamic Chat node > Voice Configuration to utilize the latest interruption handling mechanism. Ensure the Final Silence Duration is set to 0.2 seconds.

image

Improving Speech Recognition Accuracy with Boost Phrases​

Use Boost Phrases to enhance recognition accuracy for industry-specific jargon, technical terms, regional definitions, and locations. There are two ways to pass Boost Phrases:

  1. In the Question/Text Node

    • Pass a predefined list of boost phrases in the node configuration.
    • Alternatively, use a function to dynamically generate an array of boost phrases and pass them via custom voice options under the boost_phrases key. image

  2. In the Dynamic Chat (DC) Node

    • A fixed set of words and phrases can be passed via Dynamic Node > Voice Configuration > Boost Phrases. image

Knowledge Base (KB) Integration​

Knowledge Base searching is now available in VoiceX conversations via Skills. You can either:

  • Create a dedicated KB skill, or
  • Enable document search within an existing skill.

Reducing User-Perceived Latency​

To minimize user frustration during the KB search wait time, use the following techniques:

  1. Acknowledge the request with preemptive messages. These messages will play in parallel to the KB search function, reducing perceived wait time by 1-3 seconds. Examples:

    • “Hold on, let me fetch it for you.”
    • “Give me a moment to check this.”

  2. Set expectations upfront with conversational fillers. These create a smoother, more human-like interaction, reducing frustration during brief wait periods. Examples:

    • “This might take a few seconds, hang tight!”
    • “Give me some time, I’m fetching it for you.”
    • “Please wait, I’m finding the best answer.”

image

note

KB responses take 7-15 seconds to generate. Efforts are underway to reduce this latency.

info

Learn about Dynamic chat node in detail here.

Last updated on