Appearance
8.6 — The Telephone Story: From Bell to VoIP
Every design decision in modern communications was made by somebody solving a specific problem with the technology of their decade. Understanding why the telephone network is shaped the way it is explains a surprising amount of what you use today — including why a phone number looks like it does, why a call sounds the way it does, and why the internet was built the way it was.
This chapter is the origin of constraints that still shape Chapter 8.1's cellular network — the 300 to 3400 Hz voice band, the idea of a circuit held open for a call, and the numbering plan — all of which were decided by what was physically possible with the equipment described below.
1. The invention
1876. Alexander Graham Bell files his patent on 14 February; Elisha Gray files a caveat on the same principle the same day, hours apart. Bell's patent was granted, and the priority dispute has never entirely settled.
The device was extraordinarily simple. A diaphragm attached to a coil in a magnetic field. Sound moves the diaphragm, the coil moves in the field, and a current is induced. The same device in reverse reproduces the sound.
It needed no battery at all. Speaking into the transmitter generated the current that drove the receiver. That is why the earliest telephones worked over a single pair of wires with nothing else in the circuit — and why they were nearly inaudible beyond a few hundred metres.
The carbon microphone, developed by Thomas Edison and David Hughes around 1878, transformed it. Carbon granules between two plates change resistance as the diaphragm presses them, modulating a battery current rather than generating it. That gave gain — a genuine amplifier decades before the vacuum tube — and it made long-distance telephony possible.
The carbon microphone stayed in service for a century, in essentially unchanged form, until electret microphones replaced it in the 1980s.
The device in that patent is almost embarrassingly simple, and worth stating plainly because everything since is a refinement of it. Speak at a diaphragm; it vibrates; its vibration varies a current; the varying current drives a diaphragm at the far end; that diaphragm pushes air. Sound becomes electricity becomes sound. There is no amplification, no encoding and no switching in the original — just a continuous analogue of the pressure wave, which is where the word analogue comes from.
Two consequences of that simplicity shaped the next century. Because the signal was a direct copy of the sound, any noise the line picked up was heard as noise and could never be removed — which is the problem digital telephony finally solved. And because each conversation needed its own continuous electrical path, connecting everyone to everyone needed a wire per pair, which is the problem the next section is about.
2. Switching
Bell's first installations were point-to-point: a wire between two specific places. With n subscribers wanting to reach each other, that needs n(n-1)/2 wires — 4,950 for 100 subscribers, and half a million for a thousand.
The exchange solved it. Every subscriber has one pair of wires to a central point, and an operator connects any pair on request. n wires instead of n^2/2.
The Strowger switch, 1891, automated it — and the story of its invention is genuine. Almon Strowger was an undertaker in Kansas City whose competitor's wife worked as the local telephone operator, and was, he believed, diverting calls for undertakers to her husband. He built a switch to remove the operator from the process.
The mechanism is a rotating and stepping contact arm driven directly by the pulses from a rotary dial. Dial the digit 5 and the line is interrupted five times, stepping the switch five positions.
This is why the rotary dial worked the way it did, and why the digit 0 sends ten pulses — it is the tenth position, not the zeroth.
Crossbar switching, 1930s, replaced the per-line mechanism with a matrix of relays and a shared control unit. Faster, more reliable, and it separated the control logic from the switching fabric — an architectural idea that recurs everywhere in computing.
Electronic switching from 1965 put a computer in charge. This is when the network gained features: call forwarding, three-way calling, caller identification, and — commercially most significant — itemised billing.
Digital switching from the 1970s stopped switching voltages entirely and started switching timeslots in a TDM stream (Chapter 7.4). A digital exchange is a computer rearranging bytes, and the "switch" is a memory read at a chosen address.
3. Why a call sounds the way it does
The 300 Hz to 3400 Hz band was not an arbitrary choice.
The lower limit removes mains hum and rumble, and speech below 300 Hz contributes to warmth but little to intelligibility.
The upper limit was chosen because raising it costs bandwidth on every trunk circuit, and 3400 Hz gives about 90% of speech intelligibility.
What is lost is the consonants. The distinction between "s" and "f" lives mostly above 4 kHz and is essentially absent from a narrowband call. That is precisely why people spell things out on the phone and why "S for Sugar, F for Freddie" alphabets exist — the network genuinely does not carry the information needed to tell them apart.
Sampling at 8 kHz follows from Chapter 4.7: twice 3.4 kHz is 6.8 kHz, and 8 kHz leaves room for the anti-alias filter's transition band.
8 bits per sample with companding. Linear 8-bit coding would give only 48 dB of dynamic range, which is inadequate for speech that varies enormously in level. Companding applies a logarithmic curve before quantising, so quiet sounds get fine steps and loud ones get coarse steps — matching the ear's own logarithmic sensitivity.
A-law in Europe, µ-law in North America. Both give roughly the dynamic range of 12-bit linear coding in 8 bits, and the incompatibility between the two is a real nuisance at international gateways, requiring conversion that slightly degrades the audio.
8000\ \text{samples/s}\times8\ \text{bits}=64\ \text{kbit/s}
That 64 kbit/s figure became the fundamental unit of the entire global telephone network, and every trunk rate is a multiple of it: 24 channels in a 1.544 Mbit/s T1, 30 in a 2.048 Mbit/s E1, and upward through the hierarchy.
4. Signalling
In-band signalling sent control tones through the same path as the voice. Convenient, and catastrophic for security.
The 2600 Hz tone told the network that a trunk was idle. In the 1960s it was discovered that whistling or playing that tone during a call released the trunk while keeping the local connection, allowing the caller to dial a new route for free.
A toy whistle given away in Cap'n Crunch cereal produced exactly 2600 Hz, which is how John Draper acquired the name Captain Crunch. The blue box that automated this was built and sold by, among others, Steve Wozniak and Steve Jobs before they founded Apple.
Out-of-band signalling — Signalling System 7, from 1975 — fixed it by moving all control onto a separate packet network that subscribers cannot reach.
SS7 also enabled things that in-band signalling could not: caller ID, toll-free number translation, number portability, and the roaming lookups that Chapter 8.1 described. It is still in use, and its own weakness is now well documented — it was designed when only a few trusted national operators could connect to it, and that assumption no longer holds. SS7 attacks can intercept text messages, which is one reason SMS is a poor second authentication factor.
DTMF — dual-tone multi-frequency, 1963 — replaced pulse dialling. Each key sends two tones simultaneously, one from a low group and one from a high group.
| 1209 Hz | 1336 Hz | 1477 Hz | |
|---|---|---|---|
| 697 Hz | 1 | 2 | 3 |
| 770 Hz | 4 | 5 | 6 |
| 852 Hz | 7 | 8 | 9 |
| 941 Hz | * | 0 | # |
Why two tones rather than one. A single tone could be imitated by speech or by music on hold. A pair of tones from two disjoint groups, neither of which is a harmonic of the other, essentially never occurs naturally — and the frequencies were chosen specifically so that no two are in a simple harmonic ratio.
Decoding DTMF is Chapter 5.1's Goertzel algorithm, running eight filters, and that is exactly what every exchange did for forty years.
5. The trunk network
Frequency division multiplexing (Chapter 7.4) stacked analog calls onto coaxial cable: 12 channels to a group, 5 groups to a supergroup, and upward to 10,800 calls on one cable.
Digital hierarchy replaced it. E1 at 2.048 Mbit/s carrying 30 calls, then 4 E1s into an E2, and so on. The problem was that each level's multiplexer added its own justification bits to accommodate slightly different clock rates, so extracting one 64 kbit/s channel from a high-rate stream required demultiplexing every intermediate level.
SDH and SONET, from the late 1980s, fixed it by synchronising the whole network to one clock and using pointers to locate each channel. Any channel can be extracted directly — the add-drop multiplexer — which transformed network operations.
And SDH ring topologies gave 50 ms protection switching: a fibre cut anywhere on the ring reroutes traffic the other way within 50 milliseconds, which is short enough that a voice call does not drop. That reliability standard shaped what operators expected, and it is why early IP networks were considered unacceptably fragile by comparison.
6. The move to packets
Circuit switching reserves a path for the whole call. Guaranteed bandwidth, guaranteed low latency, and the resources are wasted during every pause in the conversation — which is roughly 60% of the time.
Packet switching sends addressed chunks that share the links. Far more efficient, and the delay varies because a packet may wait behind others.
Voice over IP made packet switching carry voice, and it required solving three problems that circuit switching never had.
Jitter. Packets arrive with varying delay. A jitter buffer holds arriving packets and plays them out at a constant rate, trading added latency for smoothness. Adaptive buffers grow and shrink with measured network conditions.
Packet loss. A lost packet is a gap in the audio. Packet loss concealment synthesises a plausible replacement from the preceding audio — repeating the last pitch period, with fading. Up to about 5% loss is largely inaudible; beyond that, speech becomes choppy.
Latency budget. The ITU's guidance is that one-way delay under 150 ms is transparent, 150 to 400 ms is acceptable with noticeable interruption, and above 400 ms conversation becomes difficult — people start talking over each other because the turn-taking cues arrive too late.
The codecs, and the trade each makes:
| Codec | Rate | Notes |
|---|---|---|
| G.711 | 64 kbit/s | plain PCM, no compression, universal |
| G.729 | 8 kbit/s | good quality, was patent-encumbered |
| G.722 | 48–64 kbit/s | wideband, 50 Hz–7 kHz |
| Opus | 6–510 kbit/s | adaptive, wideband, the modern default |
Opus is the one that won, and deservedly: it switches automatically between a speech mode and a music mode, adapts its rate to the network, and is royalty-free. Every browser-based call uses it.
The economics that drove the transition. A traditional long-distance call was billed per minute because the circuit was genuinely reserved. A VoIP call uses a few tens of kilobits per second of a link that is already paid for, so the marginal cost is effectively zero. The entire long-distance telephony business, worth hundreds of billions annually, was dismantled by that fact within about fifteen years.
7. Why phone numbers look like that
The E.164 format — up to 15 digits, with a country code, an area code and a subscriber number.
Area codes were assigned in 1947 for North America with a specific rule: the codes for the busiest areas had the fewest pulses to dial. New York got 212, Los Angeles 213, Chicago 312 — all short to dial on a rotary phone, because 1 is one pulse and 0 is ten. A caller in a rural area with a low-traffic code was spending a lot of finger time, and that was a deliberate optimisation for the equipment of the day.
The middle digit was originally 0 or 1, which distinguished an area code from a local exchange prefix. That restriction was abandoned in 1995 when codes ran out, which is why modern area codes look less patterned.
Emergency numbers. 999 was chosen in Britain in 1937 partly because it could be dialled in the dark by feeling for the finger stop, and partly because a long pulse train was unlikely to be produced accidentally by a faulty line. 911 was chosen in the United States in 1968 for the opposite reason — it is fast to dial. 112 is the European and GSM standard, and every GSM phone routes it to emergency services regardless of country, network or whether a SIM is present.
8. What the telephone network taught, and what it lost
The idea it got right, and which the internet adopted: universal addressing. Any number reaches any phone anywhere, through a hierarchy of routing decisions made by independent operators who agree only on the numbering plan. That is the same architectural insight as IP addressing.
The idea it got wrong: intelligence in the network. The telephone network put the features in the exchanges and made the endpoints dumb. Adding a service meant upgrading the network, which took years and required every operator to agree.
The internet inverted it — a dumb network that only forwards packets, with all the intelligence at the endpoints. Adding a service means writing software on two computers, which is why the internet accumulated capabilities at a speed the telephone network never approached.
What was genuinely lost in the transition:
Reliability. The telephone network targeted 99.999% availability — five minutes of downtime per year. It was powered from the exchange, so it worked in a power cut, and the batteries there ran for days. A VoIP phone dies with the mains, and a mobile network's cell sites have hours of backup at best.
Guaranteed latency. A circuit-switched call had constant delay by construction. A VoIP call's quality depends on network conditions, which is why calls degrade at busy times in a way that fixed-line calls never did.
Emergency location. A fixed line's address was known exactly. A mobile or VoIP call's location is estimated, and getting it right has required substantial extra machinery — which is why emergency services still ask where you are.
The honest assessment. The packet network is enormously more capable and measurably less reliable, and the trade was worth making. But the reliability of the old network was a real engineering achievement, built deliberately over a century, and it is worth knowing that it was not an accident of simpler times.
Part 9 leaves communication entirely and covers the electronics that moves power rather than information — transformers, motors, batteries and the converters that connect them.
Every formula above, built from scratch
None of the results in this chapter are worth memorising, because each one can be rebuilt in under a minute from something simpler. What follows is that rebuilding, one result at a time, so the formula and the reason for it sit on the same page as the explanation that needed them.
Telephony
\text{bit rate}=8000\ \text{samples/s}\times8\ \text{bits}=64\ \text{kbit/s}
Voice band 300–3400 Hz. Sampling at 8 kHz leaves 600 Hz for the anti-alias filter transition above 3.4 kHz.
Companding:
\text{A-law: } F(x)=\frac{A|x|}{1+\ln A}\ \text{for } |x|\lt\frac1A, \qquad A=87.6
\text{mu-law: } F(x)=\frac{\ln(1+\mu|x|)}{\ln(1+\mu)}, \qquad \mu=255
Both give roughly 12-bit linear dynamic range in 8 bits, by matching the ear's logarithmic sensitivity.
Trunk hierarchy:
\text{E1}=32\times64\ \text{kbit/s}=2.048\ \text{Mbit/s}, \qquad \text{T1}=24\times64+8=1.544\ \text{Mbit/s}
DTMF frequencies. Rows 697, 770, 852, 941 Hz; columns 1209, 1336, 1477 Hz. No two are in a simple harmonic ratio, which is what makes accidental triggering by speech or music essentially impossible.
Decoded by eight Goertzel filters (Chapter 5.1), which is cheaper than an FFT because only eight frequencies matter:
K=8 \lt \tfrac12\log_2N \ \text{for } N\gt65{,}536
— so for the short blocks used here, Goertzel wins easily.
VoIP latency budget (ITU G.114): under 150 ms one way is transparent; 150–400 ms acceptable; over 400 ms makes conversation difficult.
t_{total}=t_{codec}+t_{packetisation}+t_{network}+t_{jitter\ buffer}
Packet loss concealment keeps speech usable to about 5% loss.
Erlang B, for sizing how many circuits a group of users needs:
B=\frac{A^N/N!}{\sum_{k=0}^{N}A^k/k!}
with A the offered traffic in erlangs and N the number of circuits. This is the formula that decided how many trunks every exchange in the world was built with, and it is still used for call-centre staffing.
Fourteen worked problems next.
What the next chapter fixes
Every system in Parts 7 and 8 assumes power arrives at the socket. Part 9 covers where it comes from and how it is converted: why the grid is AC, how electricity and motion convert into each other, what is happening chemically inside a battery, and how a modern supply changes one voltage into another while wasting almost nothing.