Sizing and capacity
Last modified: 25 July 2026
Measured, not estimated
Benchmarked against the live processor on 24 July 2026, on a 4-vCPU VM:
| Configuration | Sustained throughput |
|---|---|
| 2 workers, 2 model threads | 26 utterances/s |
| 4 workers, 1 model thread | 38 utterances/s — 47% better, and what ships |
More parallel workers beat wider per-inference threading, because voice chat is many short utterances rather than a few long ones. Admission tops out at roughly eight simultaneous in-flight requests before the processor answers busy.
Those figures use a silence fixture, so transcription is unrealistically fast. The honest number is below.
How many players is that
With real speech the same VM sustains about 13 utterances per second, which is roughly 30 to 50 concurrently active speakers before it saturates and starts skipping moderation. Concurrently active means talking right now — not online, and not in a voice channel. People listen far more than they talk, so a server with a few hundred players online is usually well inside that.
Past saturation nothing breaks: requests are shed, the plugin fails open, and voice chat carries on unmoderated.
Raising the ceiling
One 4-vCPU box is not the top end, but CPU alone is not the way up. In order of effect:
- GPU transcription — a 5 to 10 times gain, and the real answer for large networks.
- Give the processor its own machine, isolated from anything else you run, so heavy speech cannot starve another service.
- A lighter English-only model for servers that only speak English.
- More vCPU, and more than one processor behind the same licence.
The per-server admission limit and rate limits exist so that one busy network cannot spend everyone else's headroom.