Sub-Second Conversational Voice with WebRTC
The Incord Team·June 2026·5 min read
Latency destroys immersion. If a voice agent takes two seconds to respond, the conversation feels broken. We built our Voice API on WebRTC to achieve true sub-second, human-like interaction.
Why Latency Matters in Voice
In text chat, users are accustomed to waiting. A spinning loading indicator that lasts for two or three seconds is standard UX for most LLM interfaces. But in a voice conversation, human psychology dictates that a pause longer than 500 milliseconds feels unnatural. At 1.5 seconds, the user assumes the agent didn't hear them, causing them to repeat themselves and completely derail the interaction.
Standard voice architectures rely on a chain of slow APIs: record audio, send via WebSocket, run Speech-to-Text, send text to LLM, wait for LLM text stream, send text to Text-to-Speech API, and finally stream audio back to the client. This sequential chain makes true conversational latency impossible.
Incord eliminates this overhead by utilizing WebRTC to establish a direct, peer-to-peer UDP connection between the client and our edge nodes. Audio is streamed bidirectionally in real-time, completely bypassing the handshake and overhead delays of WebSockets and REST APIs.
True Full-Duplex Interruption
Real human conversations aren't walkie-talkies; we talk over each other, we interrupt, and we pivot mid-sentence. If a user interrupts an AI agent, the agent must stop talking instantly and process the new context.
Because Incord's Voice API is built on WebRTC, our edge nodes detect user speech (VAD) in single-digit milliseconds. The moment the user speaks, the agent's audio stream is halted, and the LLM context window is immediately appended with the interruption event. It feels entirely organic.
Voice With Deep Context
A fast voice agent is still useless if it forgets who you are or what you were doing. Most voice APIs are isolated silos—they have no idea what the user was just looking at on your website or what they discussed with your text chatbot yesterday.
Because Incord's Voice API is deeply integrated into our Unified Intelligence Layer, your voice agents have instant, zero-latency access to the exact same Global Memory Graph as your text agents. A conversation started on a website can continue seamlessly over a phone call, with the voice agent recalling past preferences, recent actions, and live context without missing a beat.