Streaming forks off the path.
Audio to your agent over WebSocket is a fork from the Media Engine, not an extra relay in front of it. Additional streams for transcription or QA fork the same way.
Latency architecture
Single-hop telephony
for AI voice agents
Vobiz single-hop telephony sends call audio through exactly one internal relay, for sub-80ms one-way latency to your AI voice agent.
Read the docsFree pilot. We measure latency and MOS on your own calls.
Voice AI latency
Voice AI latency is a budget, and telephony spends it first
A conversational turn has roughly 500ms end to end before a pause feels unnatural. Four things spend it: speech-to-text, the model, text-to-speech, and the telephony network carrying audio both ways.
Telephony latency is set by how many relays the media crosses. At sub-80ms one-way, Vobiz leaves roughly 340ms of the budget for your stack.
Media path
How the single-hop architecture changes the media path for voice AI
Tuning can shorten a queue. It can't remove a multi-hop relay. So we built the media path with only one hop.
| The common pattern | The Vobiz path |
|---|---|
| Media crosses several internal relays | Media crosses exactly one internal relay |
| Signalling and media share a plane, so control traffic competes with audio | Control and media planes are fully decoupled |
| Every relay adds a queue and a jitter buffer | Carrier or BYOC trunk → Media Engine → your agent: one relay |
| Recording and processing sit inline, in the path | Streaming and recording fork off the path, never sit in it |
| Delay compounds hop by hop, and tuning cannot remove it | Delay is bounded by a single relay, by construction |
Off the path
Streaming, recording and control never sit in the media path
Recording forks off the path.
Starting a recording adds no inline delay to the live call. Call records and quality telemetry are written off the path too.
Control runs on its own plane.
Transfers, mid-call speech and routing changes travel as signalling, so they never compete with audio for the same path.
Bring your own carrier
Your own carrier gets the same single relay
A BYOC trunk attaches directly to the Media Engine, so calls on your carrier cross the same single internal relay as calls on the Vobiz network. Your carrier contract stays intact, and you can run Vobiz as a second route alongside it.
Measurements
Measurements and method
- Sub-80ms
- one-way telephony latency
- 4.2+
- PESQ MOS
Both figures come from Vobiz internal test calls. You can check them on your own traffic: run a pilot, or start free and pull MOS, jitter and packet loss from any call detail record by call_id.
Run a pilot
Measure it on your own traffic
Run a pilot and we'll measure latency, MOS, jitter and packet loss on your own calls, from the same 43-field call records you'd get in production.
Prefer to test it yourself? Start free, place a call, and pull MOS, jitter and packet loss from its call record by call_id. Call detail records →
FAQ
Latency questions
- The delay between a caller finishing speaking and hearing the agent reply. It is spent by speech-to-text, the model, text-to-speech and the network carrying audio both ways. A conversational turn has roughly 500ms before a pause feels unnatural.
- Once a call is set up, its audio crosses exactly one relay inside Vobiz, the Media Engine, between the phone network and your agent. The caller's own carrier is still part of the path; the single hop is what happens inside Vobiz.
- Roughly 500ms end to end. Telephony is the only part you can't fix later by swapping a vendor. At sub-80ms one-way, Vobiz leaves roughly 340ms for speech-to-text, the model and text-to-speech.
- Every relay in the media path adds a queue and a jitter buffer. Tuning can shorten them, but it can't remove the relay. The only fix is fewer relays, which is an architecture decision.
- No. Both fork off the media path at the Media Engine rather than sitting in it, so neither adds inline delay to the live call.
- On Vobiz internal test calls. You can check it on your own calls: every call detail record carries MOS, jitter and packet loss, fetchable by call_id.
- Yes. A BYOC trunk attaches directly to the Media Engine, so it crosses the same single internal relay as calls on the Vobiz network.
- In the US and India.
- No. Call setup and routing decisions travel on a separate signalling path. Once a call is set up, its audio crosses only the Media Engine and never passes through the systems that route, bill or control the call.
- Most telephony platforms were designed for human-to-human calls and added AI support later, through webhooks and a separate media stream, often routing audio through a home region. Vobiz was built for AI voice agents: streaming is part of the media layer, and audio crosses one internal relay at a Vobiz point of presence in the US or India.
See sub-80ms and 4.2+ MOS on your own calls before you move any traffic.
Request a pilot