A voice server looks free next to a game server. It is not free once you count the relay fan-out, the opus jitter buffer per listener, and the CPU that each concurrent talker costs when 60 people talk over each other in a raid call. And there is a second cost most operators never measure: the idle slot. Every reserved voice slot holds a TCP and a UDP port, a session table entry, and a heartbeat timer, whether or not anyone ever connects to it. On a shared voice host this is someone else’s problem. On dedicated metal you own every one of those numbers, so it is worth doing the arithmetic before you pick a daemon.
What each daemon actually is
Mumble is an open-source, client-server voice platform built around its own protocol (Mumble protocol over UDP with a TCP control channel), released under a BSD-style licence, with the server component commonly referred to as murmur. TeamSpeak 3 is a proprietary daemon, free for a limited slot count under a non-commercial licence, with a commercial licence required above that. Both are mature, both predate Discord’s dominance in casual communities, and both are still the default choice for organised play, milsim, racing leagues, and tournament comms where a persistent, admin-controllable voice channel tree matters more than a chat sidebar.
Neither daemon is heavy in the way a game server is heavy. No world simulation, no physics, no AI — the work is packet fan-out, mixing, and jitter buffering. That is why a 2-core VPS holds hundreds of slots. The mistake is sizing for the daemon and forgetting the game servers on the same metal.
Bandwidth: per slot, per talker, per listener
Voice bandwidth scales with talkers × listeners, not with slots. One person speaking to 40 listeners is 40 outbound streams; ten speaking at once is 400 if the daemon does not mix. TeamSpeak 3 and Mumble both mix per channel by default, so a channel’s outbound cost is roughly (listeners × mixed stream) regardless of talker count. That is the key architectural difference from game traffic: mixing collapses fan-out from quadratic to linear.
Use the following working figures for planning. They assume a 20 ms packetisation interval, RTP/UDP/IP headers of about 40–60 bytes per packet, and the codec’s target bitrate. All are reproducible from those inputs.
| Codec / mode | Bitrate | Payload + headers per stream | Per 40-listener channel (mixed) |
|---|---|---|---|
| Opus voice, 20 ms | 24–32 kbit/s | ~40–48 kbit/s | ~1.9–2.3 Mbit/s |
| Opus voice, 10 ms | 32 kbit/s | ~60–72 kbit/s | ~2.9–3.5 Mbit/s |
| Opus music, 64 kbit/s | 64 kbit/s | ~80–88 kbit/s | ~3.8–4.2 Mbit/s |
| Opus music, 96 kbit/s | 96 kbit/s | ~112–120 kbit/s | ~5.4–5.8 Mbit/s |
| PCM 48 kHz 16-bit mono | 768 kbit/s | ~792–800 kbit/s | ~38–39 Mbit/s |
The header overhead is why lowering the interval from 20 ms to 10 ms costs more than you expect: twice as many packets, twice as many 40-byte headers, so a 32 kbit/s stream becomes ~64 kbit/s on the wire. Only use 10 ms for competitive comms where 10 ms of latency genuinely changes an outcome. Never use PCM outside a point-to-point studio link.
CPU: what a concurrent talker costs
Mixing is cheap but not zero. A modern core mixes a few hundred simultaneous streams at 20 ms before showing load, and the real ceiling arrives as soft-interrupt cost on the network queue, not user-space CPU. A 2-vCPU box handles a 500-slot Mumble server without stress; TeamSpeak 3 is leaner per connection and famously runs 512 slots on one core.
The number that does matter is jitter. Voice is far less bandwidth-hungry than your game server and far more sensitive to packet loss and jitter. A 30 ms jitter buffer absorbs normal internet variance; a buffer that has to grow to 120 ms to hide a bad path adds more delay than the codec saved. If players complain about voice sounding “underwater” or cutting out while the game itself plays fine, the problem is almost always the path to the voice server, not its CPU.
Running voice alongside game servers on the same metal
This is where the decision gets real. Co-locating voice with game servers saves money and removes a network hop for players in the same region, but it exposes voice to game-server scheduling. A world save that stalls a disk queue for 400 ms does not hurt an Opus stream much — voice does not touch disk — but a CPU contention spike that delays the mix thread does. Pin the voice daemon to its own cores and let it keep its own band of the NIC queue.
| Placement | Pros | Cons | Fits |
|---|---|---|---|
| Same box as game server | Cheapest, one region, no extra hop | Shares CPU with tick loop; single point of failure | Community of <150, one or two game instances |
| Separate small instance, same DC | Isolated failure domain, still 1–2 ms away | Extra cost, second management surface | Most organised communities |
| Managed cloud voice | Zero ops | Monthly per-slot cost, no control over routing | Casual, low commitment |
Capacity math for a 200-player community
Take 200 registered players, 60% online at peak (120 concurrent), the voice server at 70% uptake (84 connected), and perhaps 30% of those talking at any instant, spread across 6–8 channels. With mixing, the outbound cost is the number of channels × listeners × mixed stream — not 84 × 84. That lands in the low single-digit Mbit/s range for outbound and a comparable inbound figure: comfortably inside a 1 Gbit/s port alongside game traffic.
The slot overhead is the sneaky one. At 500 licensed slots against 84 concurrent, headroom is fine — but if the daemon reserves a session entry and heartbeat timer per slot, and you also run a stats interface and a recording bot, background cost stays flat while game load moves. Measure both at real peak, not at 3 a.m. For why a dedicated machine changes these numbers, see the dedicated server options for game communities.
Practical setup checklist
These habits keep voice out of your incident list:
- Put voice on a different port range and, if possible, a different IP from game traffic so a DDoS mitigation rule for the game protocol does not silently drop voice.
- Set the jitter buffer to auto with an upper cap rather than a fixed value; a fixed 80 ms buffer is 80 ms of permanent delay.
- Keep the certificate and identity file safe: Mumble clients pin it, and rotating it without warning makes every client show a trust warning.
- Ban the default port; both daemons are scanned constantly for unauthenticated admin interfaces.
- Log channel joins and disconnects alongside game server logs — correlating a lag complaint with a voice re-connect is how you find a bad path.
- Test with 40 real listeners before launch, not 3.
Which should you run
Choose Mumble when you want an open protocol, self-hostable everything, and deep per-channel ACLs with no licence count to think about. Choose TeamSpeak 3 when your community already knows it and you need its permission model and music-channel behaviour. On dedicated metal either is a rounding error in CPU. The decision that changes your bill is whether it shares a box with the game servers — and whether you measured the peak.


Leave a Reply
You must be logged in to post a comment.