¿Cuánto ancho de banda necesita un call center?
Per-call bandwidth for G.711, G.729 and Opus, how to size upload and download for 10 or 100 agents, and the jitter and packet-loss checks that matter more.
Respuesta rápida
One call needs about 87 kbps each way with G.711 and 31 kbps with G.729 on Ethernet at 20 ms packets (Cisco). RFC 6716 puts Opus wideband speech at 16 to 20 kbit/s, before headers. Multiply the per-call rate by peak concurrent calls and add headroom: 100 agents on G.711 need roughly 11 Mbps each way in our example. Jitter and packet loss matter more than raw speed.
Search for the bandwidth a call center needs and you will find answers from 80 kbps per agent to several megabits, often with no codec and no working. The numbers are not wrong so much as incomplete: each one assumes a different codec, a different packet interval and a different amount of safety margin.
This guide starts from published per-call figures, shows the formula, works it for 10, 50 and 100 agents with clearly labelled example inputs, and then covers the checks that decide call quality far more often than a speed test does.
How much bandwidth does one call use?
A voice call is a stream of small packets, 50 a second at the usual 20 ms interval. The codec sets the size of the voice payload. Every packet then carries headers: 12 bytes of RTP, 8 of UDP, 20 of IP and, on Ethernet, 18 bytes of frame overhead. Cisco's per-call bandwidth document works the sums for two common codecs at 20 ms:
| Codec | Codec bit rate | Voice payload per packet | Packets per second | Bandwidth on Ethernet, per direction |
|---|---|---|---|---|
| G.711 (PCMU or PCMA) | 64 kbps | 160 bytes | 50 | 87.2 kbps |
| G.729 | 8 kbps | 20 bytes | 50 | 31.2 kbps |
The formula behind both rows is simple: (18 + 40 + payload bytes) x 8 bits x 50 packets a second. For G.711 that is (18 + 40 + 160) x 8 x 50 = 87,200 bits per second. The headers cost more than the voice does at low bit rates, which is why G.729 at 8 kbps still needs 31.2 kbps on the wire.
What about Opus and WebRTC?
Opus is different because it is variable. RFC 6716 says it supports all bit rates from 6 kbit/s to 510 kbit/s and gives the sweet spots for speech at 20 ms frames: 8 to 12 kbit/s for narrowband and 16 to 20 kbit/s for wideband. Opus can also encode frames from 2.5 ms to 60 ms, so the packet rate can change too.
Browser softphones matter here. RFC 7874 says WebRTC endpoints are required to implement Opus and also G.711 (PCMA and PCMU), so a browser agent desktop can use either. Which one a call actually uses is negotiated when the call is set up, so ask your platform which codecs it offers to the browser and read the negotiated codec from a test call. For a connected PBX, DialerBee's deployment page lists Opus first by default, falling back through G.722 and G.711, and you can set your own codec order. If you want the background on codec choice, What Is a Voice Codec explains the trade-offs, and SIP 488 Not Acceptable Here shows what happens when two sides share no codec.
For a rough Opus figure, apply the same header arithmetic. This is our own illustrative calculation using Cisco's overhead values, not a published number:
| Opus setting (20 ms) | Payload per packet | Illustrative Ethernet bandwidth |
|---|---|---|
| 8 kbit/s (narrowband sweet spot, low end) | 20 bytes | 31.2 kbps |
| 20 kbit/s (wideband sweet spot, high end) | 50 bytes | 43.2 kbps |
Real traffic runs a little higher. Encryption, VPN tunnels and Wi-Fi framing all add bytes, so treat the Ethernet figure as a floor and leave headroom, as the next sections do.
Why does upload fail first on office and home connections?
A call sends and receives about the same amount, so each direction needs the full per-call rate. The problem is that most consumer and small-business broadband sells a large download and a small upload. A line advertised as 100 Mbps down may offer only a fraction of that up. Your agent's voice goes up, so the small number is the one that runs out.
Three things then compete for that small upload: the call itself, anything else on the line (video meetings, cloud backups, photo sync, a family member streaming) and the extra bytes of whatever tunnel or VPN the agent uses. Voice is small, but it is also unforgiving. A big upload that fills the link makes voice packets wait in the router's queue, and that queueing shows up as jitter and choppy audio, not as a slow page.
The practical rule: look at the upload figure on the contract, not the headline speed, and test it at the hours your agents work.
How do you size bandwidth for 10, 50 or 100 agents?
Use one formula:
Bandwidth needed (each way) = peak concurrent calls x per-call rate x headroom
- Peak concurrent calls. The most agents you expect to be talking at the same moment. For a busy outbound floor, assume all of them.
- Per-call rate. From the table above for your codec.
- Headroom. The extra margin for overhead, other traffic and bursts. The 30% used below is our planning choice, not a standard. Pick your own and write it down.
Worked example (illustrative inputs)
These figures assume every agent is on a call at once, G.711 at 87.2 kbps or G.729 at 31.2 kbps, and 30% headroom. They are example inputs, not measurements from any customer.
| Agents on a call | G.711 raw | G.711 with 30% headroom | G.729 raw | G.729 with 30% headroom |
|---|---|---|---|---|
| 10 | 872 kbps | about 1.1 Mbps | 312 kbps | about 0.4 Mbps |
| 50 | 4.36 Mbps | about 5.7 Mbps | 1.56 Mbps | about 2.0 Mbps |
| 100 | 8.72 Mbps | about 11.3 Mbps | 3.12 Mbps | about 4.1 Mbps |
Each figure applies to upload and to download separately. If a supervisor listens to a live call, that is one more audio stream to count.
Which legs of the call cross your internet link?
This is where many calculators go wrong. A carrier trunk does not have to end at your office. On DialerBee, agents use a browser softphone, and recording is captured at the platform before the call goes out to the gateway, so when the trunk connects to the platform, the office link carries each agent's own call leg between the browser and the platform. The deployment page describes the three patterns, and the Traiga su propio operador page explains how your own trunk connects.
Check your pattern before you size anything:
- Voice supplied for you (DialerBee's carriers, through our sales team), or your own carrier trunk connected to the platform. Count one leg per agent call on the office link.
- Your own PBX or SBC at the office. If media also flows between the platform and a PBX that sits on your site, that second leg crosses your link as well. Count both legs as a conservative upper bound, and ask your account team to confirm the media path for your setup.
- PBX in a data centre. Its leg to the platform does not touch the office line, so the agent leg is the one to size.
Remote agents are simpler: each home line carries one call leg for that one agent, so the sizing question is per home, not per floor.
Why do jitter, packet loss and latency matter more than raw speed?
Once the link has room for the calls, extra speed adds almost nothing. What decides whether the call sounds good is how the packets arrive:
- Jitter is variation in the gap between packets. Voice needs an even stream, and a large swing makes audio choppy or robotic. What Is Jitter explains how the receiving side buffers against it.
- Packet loss means some packets never arrive, so syllables go missing. What Is Packet Loss covers causes and fixes.
- Estado latente is the delay from speaker to listener. High delay makes people talk over each other. What Is VoIP Latency goes through it.
A speed test measures how much data a link can move over a few seconds. It says little about whether a 20 ms voice packet arrives on time every time. That is why a fast home line with a busy Wi-Fi channel often sounds worse than a modest wired line. For targets and tools, see How to Measure VoIP Call Quality, and for browser-specific faults see WebRTC Agent Audio Problems.
Home agents: a 5-step connection check before the first shift
- Use a cable. Plug the computer into the router. If that is impossible, sit near the access point.
- Read the contract upload speed. Compare it with the per-call rate above. One call needs well under 1 Mbps, but the line has to share it with everything else.
- Test during the working hours. Run a speed and latency test at the time the shift will run, when neighbours and family are also online.
- Quiet the line. Pause cloud backups and large uploads, and close video streams on other devices while on a call.
- Place a real test call. Make a call of a few minutes and listen for choppy audio, echo and one-way silence. Then check the quality figures your platform shows.
For the wider picture of running a remote floor, see virtual call center software, and for what the browser softphone does, see the Escritorio del agente.
Office floor: QoS, a separate voice VLAN and a second internet line
On a shared office network, three steps protect voice:
- Quality of service (QoS). Configure the router so voice packets jump the queue ahead of bulk traffic. Without it, one large upload can delay every call.
- A separate voice VLAN. Keeping call traffic on its own network segment stops guest devices and file shares from competing with it, and makes problems easier to trace.
- A second internet line. With two lines from different providers, one failure does not silence the whole floor. Test the failover by unplugging the primary, as you would test a backup trunk.
Also keep Wi-Fi off the voice path where you can, and give any wireless area enough access points for the number of people in it.
How to test the link before go-live
- Record your inputs. Codec, peak concurrent calls, headroom, and which legs cross the link.
- Measure the baseline. Capture upload, download, latency, jitter and packet loss with nobody on a call.
- Load it. Place test calls from a few agents, then add more until you reach your expected peak. Watch the quality figures as the count rises.
- Stress it. Repeat while a large upload runs on the same network, to see whether QoS does its job.
- Check the codec. Confirm which codec the test calls negotiated, so your sizing matches reality.
- Size trunk channels separately. Bandwidth is one question and the number of simultaneous channels on your carrier is another. See How Many SIP Channels Do I Need for that calculation.
Preguntas frecuentes
How much bandwidth does one VoIP call use?
It depends on the codec and the packet interval. At 20 ms packets on Ethernet, Cisco lists 87.2 kbps for G.711 (a 64 kbps codec) and 31.2 kbps for G.729 (an 8 kbps codec), in each direction. Opus is variable: it runs from 6 to 510 kbit/s, and RFC 6716 gives 8 to 12 kbit/s for narrowband speech and 16 to 20 kbit/s for wideband speech as the sweet spot. Add packet headers to any codec rate to get the real figure.
How much upload speed does a call center agent need?
Plan for roughly the per-call rate of your codec, times one call per agent, plus headroom. With G.711 that is about 87 kbps up and 87 kbps down per agent on a call, or about 113 kbps each way with the 30% headroom used in this guide, so 150 kbps leaves some extra room. A shared home line also carries video calls, cloud backups and other devices, so test the upload speed during the busiest hour, not only at night.
Is 10 Mbps enough for a 50-agent call center?
Often yes for the voice itself. With the example inputs in this guide (G.711, every agent on a call at once, 30% headroom), 50 agents need about 5.7 Mbps in each direction. That leaves little room for anything else on a 10 Mbps upload, so check your own figures: the codec, your peak concurrency, other traffic on the link, and whether your PBX sits at the office, which adds a second leg per call.
Does Wi-Fi work for call center agents?
It can, but a cable to the router removes a common source of jitter and packet loss. Wi-Fi shares airtime with every other device nearby, and voice packets that arrive late are as bad as packets that never arrive. If agents must use Wi-Fi, keep them close to the access point, avoid crowded channels where you can, and run the same connection test described above at the time of day they work.
Which matters more for call quality, bandwidth or latency?
For a call that already fits in the link, jitter, packet loss and latency matter more than extra speed. A faster line does not fix a congested Wi-Fi channel or a router that queues voice behind a large upload. Measure all three on the path the agent actually uses, and read How to Measure VoIP Call Quality for the targets and the tools.
Do I need a dedicated line for a call center?
Not always. A small team on a stable business connection with room to spare can run voice on the same line. A larger floor benefits from quality of service rules, a separate voice VLAN and a second internet line as backup, so one cut or one saturated link does not stop every agent at once.
Fuentes
- Cisco, Voice Over IP - Per Call Bandwidth Consumption (G.711 87.2 kbps and G.729 31.2 kbps on Ethernet at 20 ms and 50 packets per second). cisco.com (checked 2026-11-19)
- RFC 6716, Definition of the Opus Audio Codec (6 to 510 kbit/s; sweet spots of 8 to 12 kbit/s narrowband and 16 to 20 kbit/s wideband at 20 ms). rfc-editor.org/rfc/rfc6716 (checked 2026-11-19)
- RFC 7874, WebRTC Audio Codec and Processing Requirements (endpoints are required to implement Opus and G.711). rfc-editor.org/rfc/rfc7874 (checked 2026-11-19)
- DialerBee deployment patterns (browser softphone, recording captured at the platform, default codec order for a connected PBX). dialerbee.com/deployment (checked 2026-11-19)
Lecturas relacionadas
- How to Measure VoIP Call Quality
- WebRTC Agent Audio Problems
- Software de call center virtual
- Escritorio del agente
- Traiga su propio operador
- Patrones de despliegue
- What Is a Voice Codec
Lecturas relacionadas: Call Center Software vs a Business Phone System
Artículos relacionados
99.9% vs 99.95% Uptime: What an SLA Covers
8 minutos de lectura
TecnologíaAI Voice Agent Latency: What Callers Notice
9 minutos de lectura
OperacionesCall Center Business Continuity Plan Template
Lectura de 11 minutos
GuíasWhat to Test in a Call Center Software Trial
9 minutos de lectura
¿Listo para ver DialerBee en acción?
Solicite una demostración en vivo de 15 minutos o comience una prueba gratuita de 14 días: sin presentaciones, sin compromiso. Las llamadas empiezan en cuanto su propia troncal SIP o PBX está conectada y aprobada.
Prueba gratuita de 14 días · crédito de IA gratuito · sin tarjeta · 11 idiomas · BYOC · controles que apoyan el cumplimiento