Latency
RTP timestamp echo. A monotonic clock embedded in the audio payload, so the figure is a true end to end audio path measurement rather than a signalling round trip. Median and P95 across 500 samples per run.
Xebit holds 9,200 concurrent one to one voice sessions per node against 2,800 for LiveKit (open source) on identical hardware, a 3.3x difference. P95 mouth to ear latency is 118 ms against 142 ms. On participants in a single room we are lower, 82 against 780. Every figure below comes with the method that produced it, and we will set you up to reproduce all of it on your own machines.
Updated
| Metric | Xebit | LiveKit | Difference |
|---|---|---|---|
| Concurrent one to one sessions per node | 9,200 | 2,800 | 3.3x |
| Participants per node, all one to one | 18,400 | 5,600 | 3.3x |
| Agent Server sessions, dedicated node | 10,500 | 18 | see note |
| Median join to audio, cold join | 191 ms | 247 ms | −23% |
| P95 mouth to ear, two party call | 118 ms | 142 ms | −17% |
| SIP call setup, PSTN ingress | 430 ms | 580 ms | −26% |
| Reconnect after forced node kill | 1.3 s | 2.2 s | −41% |
| Audio MOS at 8% packet loss | 3.9 | 3.6 | higher is better |
| Participants in a single room | 82 | 780 | we are lower |
A conferencing SFU is tuned for large rooms with a few active speakers. A voice agent platform is thousands of isolated two party sessions where someone is talking in every single one. Those are opposite optimisation targets in the media path and in per session state, and you cannot have both.
We chose the second, which is why we lead the one to one rows by roughly three times and lose the single room row by roughly nine. If your use case is a 780 person town hall, use something else. If it is voice agents at volume, this trade is the entire point.
RTP timestamp echo. A monotonic clock embedded in the audio payload, so the figure is a true end to end audio path measurement rather than a signalling round trip. Median and P95 across 500 samples per run.
Injected with tc and netem at 2 to 10 percent random loss on the sender path. Scored with a PESQ derived MOS on the recorded output, so the score comes from the received audio and not from a loss counter.
Measured from SIP INVITE at a real carrier test trunk to the first inbound RTP packet observed at the Bridge, averaged over 100 calls. A real trunk, not a loopback.
An active call is sustained while the serving SFU process is forcibly terminated. Reconnect is measured from the disconnect event to audio resumption at the client SDK.
The harness ramps concurrent isolated two party calls rather than growing a single room. Participant count is reported as twice the maximum stable session count.
Every test repeated five times and reported as a trimmed mean, with the top and bottom runs excluded. Metrics collected with Beats on each node and shipped to an ELK stack, so instrumentation lives outside the system under test.
In April 2026 we ran a national political campaign in India: 2.5 million calls in 20 days, 150,000 a day sustained for the full month, at 67% pickup and engagement, on a single VM. The compute bill for that month was about $800. It was a one time project.