Free tool
Voice latency test
Measure what this connection and this device add to a spoken turn - network round trip, WebRTC stack overhead, and the audio latency your browser reports.
The microphone step is optional and asks for permission. No audio is recorded, transmitted or stored - the only thing read is how long the device takes to hand over a stream and what latency it reports.
What this measures - and what it cannot
Being precise about this matters more than the numbers themselves.
| Measurement | What it really is |
|---|---|
| Network RTT | Round trip from your browser to this server, over HTTPS. A good proxy for your connection quality, and not the path a conversation takes - that goes to a LiveKit media server whose address only appears in a real session's wsUrl. |
| Jitter | Mean deviation between samples. More predictive of bad audio than raw latency: a steady 120 ms sounds fine, a bouncing 40–200 ms does not. |
| WebRTC loopback | A data channel between two peer connections inside this tab. It proves WebRTC works here and shows the browser's own stack overhead. It crosses no network, so it is a floor, not a forecast. |
| Mic acquisition | Time from requesting the microphone to holding a live track. Slow here means a slow start to every conversation on this device. |
| Audio I/O latency | What the browser reports for its own audio pipeline (baseLatency plus outputLatency). Bluetooth headsets are usually the reason this is large. |
None of this measures how fast the model answers. It measures the parts around it - the parts you can actually change. If the numbers here are good and calls still feel slow, the next suspect is a tool call, and that one is yours.
Reading the numbers
| Metric | Comfortable | Noticeable | Problem |
|---|---|---|---|
| Network RTT median | < 60 ms | 60–150 ms | > 150 ms |
| RTT p95 | < 90 ms | 90–220 ms | > 220 ms |
| Jitter | < 15 ms | 15–40 ms | > 40 ms |
| Mic acquisition | < 300 ms | 0.3–1.2 s | > 1.2 s |
For context: human turn-taking gaps sit in the low hundreds of milliseconds. Around a second reads as hesitation. Past two seconds, callers start talking over the agent. Everything in the chain spends against that - the voice page breaks the rest of it down.
If the numbers are bad
- High jitter, decent median - usually Wi-Fi. Test again on a cable before changing anything else.
- High median everywhere - distance or a VPN. A corporate VPN routing through another continent is the classic cause.
- WebRTC blocked - a firewall or extension. Voice will not work at all in this browser; the text mode still will.
- Slow mic acquisition - a device problem, not a network one. Bluetooth pairing and virtual audio devices are the usual suspects.
- Large audio I/O latency - try wired headphones. Bluetooth adds delay in both directions.