Endpointing

In voice AI, endpointing is the process of detecting when a speaker has finished talking, so the AI knows when to start responding. It's controlled by a silence threshold measured in milliseconds.

Technical definition

Endpointing (also called end-of-utterance detection or voice activity detection) uses a combination of silence duration analysis and acoustic modeling to determine when a speaker's turn is complete. The silence threshold parameter controls how long the system waits after the last detected speech before considering the utterance complete. Lower thresholds (e.g. 300ms) create faster but potentially premature responses; higher thresholds (e.g. 700ms) feel more natural but add latency.

How Verabase uses Endpointing

In Verabase, endpointing is configured per agent under Agent Settings → Voice → Endpointing. The default silence threshold is 500ms — a balanced setting for most use cases. For fast-paced interactions (like order lookups), you can lower it to 300ms. For consultative conversations (like technical support), 700ms prevents the AI from cutting off users who pause to think. Verabase uses Deepgram's Nova-3 model for real-time endpointing with sub-100ms detection latency.

Why it matters for support teams

Endpointing directly affects how natural a voice AI conversation feels. Too aggressive (low threshold), and the AI interrupts the caller mid-sentence. Too passive (high threshold), and there are awkward silences after every question. Getting endpointing right is the difference between a voice agent that feels robotic and one that feels conversational.

Related terms

How Verabase Handles Self-Healing Knowledge

The standard industry approach relies on support agents manually flagging stale content and technical writers updating documentation weeks later. Verabase completely upends this model through an autonomous AI platform that identifies knowledge gaps the moment an AI agent encounters a question it cannot answer.

Rather than requiring manual intervention, our visual RAG engine drafts a proposed fix. It analyzes the context of the user's question, reviews your existing knowledge base for conflicts, and creates a clear, structured article or snippet. All you need to do is click 'Approve'. This ensures your AI agents continually get smarter over time without adding to your support team's workload.

Security First Architecture

Many legacy platforms bolt AI onto their existing infrastructure, creating potential data leaks between tenants or exposing sensitive internal documents to end-users. Verabase is built from the ground up with a Bring Your Own Key (BYOK) architecture, enterprise-grade access controls, and strict semantic boundaries. This means your private engineering documents never accidentally leak into customer-facing agent responses.