The most common bandwidth mistake in call centers is buying more of it to fix a problem that is not bandwidth.

How much you actually need

Voice is a remarkably light application. The numbers surprise people.

CodecPer call, per direction10 agents50 agents
G.711 (ulaw/alaw)~87 kbps0.9 Mbps4.4 Mbps
G.729~31 kbps0.4 Mbps1.6 Mbps
Opus (typical)~40 kbps0.4 Mbps2.0 Mbps

Fifty agents on G.711 need under 5 Mbps in each direction. A modern connection has this many times over.

Where the overhead comes from

G.711 encodes 64 kbps of audio. The wire figure is 87 kbps because every 20 ms packet carries:

Audio payload      160 bytes
RTP header          12 bytes
UDP header           8 bytes
IP header           20 bytes
Ethernet framing    18 bytes
------------------------------
Total              218 bytes  ×  50 packets/sec  ×  8  =  87.2 kbps

Smaller packets mean more overhead. Some operators use 30 ms packetisation to reduce it, at a small cost in latency.

Size for peak plus headroom: take your maximum concurrent calls, add 50%, and leave the rest of the connection for everything else the office does.

The three numbers that actually matter

MetricGoodAcceptableProblem
Latency (one-way)< 100 ms100-150 ms> 150 ms
Jitter< 20 ms20-30 ms> 30 ms
Packet loss0%< 1%> 1%

Latency is delay. Beyond about 150 ms one-way (ITU-T G.114), people start talking over each other because the natural conversational rhythm breaks. It is the metric agents describe as "the customer keeps interrupting me."

Jitter is variation in packet arrival. Voice needs packets at even intervals; when they arrive irregularly, the jitter buffer either waits (adding latency) or discards them (creating gaps). Jitter produces the robotic, choppy sound.

Packet loss is unrecoverable. TCP retransmits; voice cannot — a packet arriving late is useless. Loss becomes missing syllables.

Testing properly

Speed tests are close to useless for this. They measure throughput on a single TCP stream to a nearby server. You need loss and jitter over time, to the place your calls actually go.

Continuous ping

The simplest useful test — run for at least an hour during peak:

bash
ping -i 0.2 -c 3000 sip.yourprovider.com | tail -3

Look at min/avg/max/mdev. mdev is your jitter. A max far above avg means periodic congestion.

MTR — where the problem is

bash
mtr --report --report-cycles 100 sip.yourprovider.com

MTR shows loss at each hop. Loss appearing at one hop and continuing to the destination is real. Loss at a single intermediate hop that does not persist downstream is usually just ICMP de-prioritisation, not a genuine fault — a distinction that saves many pointless support tickets.

Under load

Test while the connection is busy. Start a large download, then run the ping test. A connection that is perfect when idle and terrible under load has a bufferbloat problem, fixed with queue management (fq_codel or cake) on the router, not with more bandwidth.

Fixing quality problems

Prioritise voice with QoS

On the office router, mark RTP traffic (DSCP EF, value 46) as highest priority. Without this, one agent's large file download can degrade every call in the building.

Wire everything

Wi-Fi is the single largest source of unexplained call quality complaints. Interference, roaming between access points, and shared airtime all produce exactly the symptoms above, intermittently, which makes them maddening to diagnose. Run Ethernet.

Turn off SIP ALG

Consumer and small-business routers frequently include "SIP ALG," which attempts to rewrite SIP messages passing through NAT. It almost always corrupts them. Disable it. This one setting is responsible for a remarkable share of VoIP failures.

Separate voice and data

If budget allows, a dedicated connection for voice removes the entire class of "someone started a backup and calls broke."

Check the server

Choppy audio can be server-side. On the dialer:

bash
uptime                      # load average vs core count
free -m                     # swap in use is a red flag
iostat -x 1 5               # %util near 100 means disk saturation
asterisk -rx "core show channels"

Call recording writes constantly. On an undersized VPS with slow storage, disk I/O saturation produces audio artefacts that look exactly like network problems.

Diagnostic table

SymptomLikely causeCheck
One-way audioNAT or RTP ports blockedexternip, firewall UDP 10000-20000
No audio either wayCodec mismatchSIP 488; compare allow= lines
Choppy / roboticJitterping mdev; Wi-Fi; QoS
Missing syllablesPacket lossmtr
EchoAcoustic feedbackHeadset, not speakerphone; check gain
Talking over each otherLatency > 150 msping avg; carrier route
Audio dies after ~30 sNAT keepalive / firewall timeoutqualify=yes, session timers
Fine when quiet, bad at peakCongestion or bufferbloatTest under load
Fine on some destinations onlyCarrier route qualityTest alternate route
Crackling on one agent onlyHeadset or USB portSwap hardware first

Home-based agent minimums

For remote agents, set a floor and test it before they start:

  • 10 Mbps down / 3 Mbps up minimum, wired
  • Latency to your server under 100 ms
  • Jitter under 20 ms, loss under 0.5%, measured over an hour at their normal working time
  • No Wi-Fi, no mobile hotspot, no shared VPN
  • UPS if local power is unreliable

Have every remote candidate run a one-hour ping test and send you the output before their first shift. It takes them ten minutes and saves weeks of ambiguous quality disputes.

Where to go next

Frequently asked questions

How much bandwidth does one VoIP call use?

About 87 kbps in each direction using G.711 with 20 ms packets, once RTP, UDP, IP and Ethernet headers are counted. G.729 reduces this to roughly 31 kbps per direction. Ten simultaneous agents on G.711 therefore need under 1 Mbps each way, which is far less than most people expect.

Why do my calls sound choppy on a fast connection?

Choppy audio is almost always jitter or packet loss, not insufficient bandwidth. A 500 Mbps connection with 60 ms of jitter sounds worse than a stable 10 Mbps line. Common causes are Wi-Fi, an overloaded router, ISP congestion at peak hours, or SIP ALG corrupting packets.

Should call center agents use Wi-Fi?

No. Wi-Fi introduces variable latency and packet loss that degrade voice quality unpredictably, and it is the most common cause of complaints that appear random. Run wired Ethernet to every agent position. It is the cheapest quality improvement available.

What is an acceptable packet loss rate for VoIP?

Below 1 percent is acceptable and generally imperceptible. Between 1 and 3 percent produces noticeable dropouts. Above 3 percent makes conversation difficult. Unlike file transfers, voice cannot retransmit lost packets, so loss translates directly into missing audio.