How one spoken “hello” crosses a cell network
You say “hello.” About a tenth of a second later your friend hears it, maybe on the other side of the country. In between, your voice becomes numbers, then a radio signal, then light in a glass fiber, then a radio signal again, then sound. Click any part of the picture to open it up.
How long does it take?
The time from your mouth to your friend's ear. Phone engineers aim to keep it under 150 milliseconds, the point where a conversation starts to feel delayed (ITU-T recommendation G.114).
What gets simplified here
- The page follows a 4G LTE voice call (“VoLTE”). 5G works the same way at this level of detail.
- Your voice is not sent as one packet. It leaves as a packet every 20 milliseconds, 50 per second, each carrying a slice of sound.
- Both towers may belong to one carrier, or the call may cross between carriers. The idea is the same.
- Numbers are typical values, so a real call will differ.
Joining the network, and never stopping looking
From the moment it is switched on, a phone finds a tower, gets on the air, proves who it is, and registers. Then, for the rest of the day, it stays reachable while sleeping most of the time, and keeps checking whether another tower would be better. Step through its life.
A day in the life of a phone
Not to scale: a phone is idle more than 95% of the day. Who looks for a better tower changes with the state. While idle, the phone decides and moves on its own. While connected, the network decides, from the phone's reports. Click an event to jump to it.
Before the first “hello”: setting up the call
Both phones have already joined the network. When you dial, the network has to find your friend's phone, wake it up, and reserve priority for the call. It takes a second or two, using control messages that never carry sound. Step through it.
Inside the phone: from sound to radio waves
Your phone runs the same assembly line 50 times a second, once for every 20 milliseconds of speech. Click a stage. The orange number is how many bits that stage hands on for one 20 ms slice of “hello.”
The radio link: sharing the air with everyone else
Hundreds of phones near one tower all want to send at once, on the same few frequencies. The tower acts like a traffic controller. Every millisecond it decides which phone may use which slice of frequency, and tells them.
Your 20 ms of “hello,” broken down for the air
Reading the grid
Each row is one resource block of frequency (12 subcarriers, 180 kHz). Each column is one millisecond. Time moves left. This is a 5 MHz channel, so there are 25 rows.
Your phone gets 4 blocks once every 20 ms, just enough for one voice packet, then stays quiet. Voice is scheduled first because it cannot wait. A video call next to you gets whatever is left, which is most of the grid.
Sometimes the tower cannot decode a packet. Then it asks again and your phone resends it 8 ms later (striped blocks). This is HARQ.
The top and bottom rows carry control messages: acknowledgments, “I have data to send” requests, and signal-quality reports.
The older idea: take turns in time
2G (GSM) phones shared each 200 kHz frequency by taking turns. Time is cut into frames of 4.615 ms, and each frame into 8 slots of 0.577 ms. Your phone was given one slot, slot 3 here, and transmitted only then. It stayed silent for the other 7, every frame, for the whole call.
That is the “rotating pattern” you can watch on the left. It is simple, but a slot is reserved even when you are silent, and one phone can never use more than its slot.
4G and 5G keep the idea of slots in time and add slots in frequency, assigned fresh every millisecond. Nothing is reserved, so the space you are not using goes to someone else.
Up and down are separate
Your phone talks to the tower (uplink) on one band of frequencies and listens (downlink) on another, so both directions work at the same time. In the example band, the phone sends at 777 to 787 MHz and listens at 746 to 756 MHz. This is called FDD. Some 5G bands use one frequency and alternate in time instead (TDD).
The cell tower
A tower is a radio station, a computer and a network connection on a pole. It decodes your signal, decides who transmits when, and hands your packets to the carrier's fiber network. Click the parts of the tower.
Handover: keeping the call going while you drive
Your phone keeps measuring the signal from nearby towers and reports it. When a neighbor is clearly stronger, at least 3 dB for a moment, the network moves your call to it. You hear nothing, because the switch takes about 50 ms and packets in flight are forwarded to the new tower.
Roaming: using someone else's towers
The carrier network: fiber, routers and the mobile core
Past the tower, your voice is ordinary internet packets inside the carrier's private network. Two kinds of traffic travel here: the voice packets themselves, and the control messages that set up the call and track where every phone is. Click a router, a fiber, or a box.
Hop by hop, Ohio to California
Light in glass fiber travels about 200,000 km per second, two-thirds of its speed in a vacuum: 5 microseconds per kilometer. Each router adds a few tens of microseconds unless its queues are full.
Inside a router: look up, queue, send
Every packet that reaches a router gets the same three steps. Read the destination address, look it up in the routing table to choose an output port, then wait in that port's queue until the link is free. Turn up the traffic and watch what happens to the queues and to your voice packets.
Step 1: the routing table
Each entry is a prefix: an address pattern where only the first bits must match. 10.20.0.0/16 means “any address whose first 16 bits are 10.20.” When several entries match, the router picks the longest one, the most specific. Type an address to try it.
Routers build this table automatically by trading “I can reach these addresses” messages with their neighbors, using protocols such as OSPF and BGP. A backbone router holds about a million entries and looks one up in nanoseconds with special memory (TCAM).
Steps 2 and 3: queues
An output port can send only so fast. When packets for one port arrive faster than that, they wait in a queue. Waiting adds delay, and a full queue has no choice but to drop new packets.
- Above 100% load the queue on port 2 can only grow. Drops are inevitable.
- A bigger buffer drops less but makes everyone wait longer. Engineers call too much buffering “bufferbloat.”
- With voice priority on, voice packets get their own queue that is always served first (the carrier marks them “expedited” in the IP header). Turn it off at high load and see what happens to the voice delay.
The simulation uses a 100 Mbit/s port, so one 125-byte cell takes 10 µs to send. A dropped voice packet is 20 ms of missing sound, which the receiving phone tries to hide.
The receiving phone: from radio waves back to sound
Your friend's phone runs the phone's assembly line backwards. Two jobs are new on this side: guessing the right symbols through radio noise, and smoothing out packets that arrive at uneven times.
Steps 2 and 3: noise, guesses and error correction
Each small dot is one received symbol, pushed off its ideal spot by noise. The phone decides by picking the nearest ideal point, so dots that drift across a boundary (red) become wrong bits. Lower the signal quality, or switch to 16-QAM, whose points are closer together.
The error-correcting code repairs a surprising number of wrong bits, then fails quickly past a limit. That is why calls sound perfect until they suddenly break up. If the CRC still fails, the tower resends, up to 3 times.
“Packet decoded” uses an approximate error curve for the rate-1/3 turbo code. SNR is the signal power compared with the noise power: 0 dB means equal, 10 dB means the signal is 10 times stronger.
Step 6: the jitter buffer
Packets leave exactly 20 ms apart but cross the network in slightly different times, because of queues, resends and handovers. If the phone played each one the moment it arrived, the voice would stutter. Instead it waits a little and plays them on a steady clock. Any packet that arrives after its turn is too late, and counts as lost.
Steps 7 and 8: back to sound
The decoder rebuilds each 20 ms slice from its description. For a missing slice it guesses, usually by repeating the last good slice and fading it. That is packet loss concealment. One missing slice is hard to notice. Several in a row sound like a robot voice or a dropout.