PSTN
In one sentence
The PSTN, or Public Switched Telephone Network, is the worldwide arrangement of carriers, numbering plans and interconnect deals that lets any phone number dial any other, and the narrowband audio it carries holds measurably less information than a browser microphone captures.
Not to be confused with WebRTC, or SIP and SIP trunking.
Definition
The PSTN is the ordinary telephone network, the worldwide arrangement that lets any phone number dial any other. Reach a normal number and you have touched it.
Its job was to move human speech on as little bandwidth as possible, so it throws away whatever it can spare. Listeners barely register the loss. Speech recognition software does.
Public Switched Telephone Network is the full name, and the thing it names is an arrangement rather than an object: carriers, national numbering plans, interconnect deals, and the regulation wrapped around all three. Internally it went digital and IP long ago. What it never shed were the limits it was born with, and those limits are the whole story for voice AI.
The audio constraint that shapes everything
- A phone channel is narrowband. Convention puts it at roughly 300 to 3,400 Hz, sampled at 8 kHz. That was a deliberate call about how little bandwidth intelligible speech needs, taken long before anything we would now recognize as digital audio existed.
- Recognizers score measurably worse on it than on wideband browser audio, and the reason is not the model. Several consonant distinctions sit in frequencies the network drops, so the evidence a recognizer would need never turns up.
- That is why a single agent behaves one way on a call and another way on a site, and why a word error rate measured on clean wideband material flatters what the same system will do over a phone line.
The regulatory and commercial constraints
- National numbering plans hand out numbers. They are regulated assets, not strings anybody can mint.
- Interconnect runs on money. Starting and finishing a call both cost something, and what they cost depends on where the call lands.
- Reputation sticks to a number, which is exactly why spam labeling and caller authentication became problems worth solving.
How voice AI touches it
- Give an agent a phone number and it gets there over a SIP trunk or through a CPaaS provider, picking up every limit listed above on the way in.
- Run the same agent in a browser and none of it applies, because the call never enters the phone network.
The direction of travel
- Plenty of markets are switching off circuit-switched kit and moving to IP interconnect, and some paths will carry a wideband codec end to end.
- Send a call across a few carriers, though, and it usually transcodes down anyway. The weakest link in the chain sets the quality, which is why nobody should plan around wideband on a phone call.
Common misconception
That call quality stopped being an issue years ago. Hand a call between carriers and it still lands on narrowband, and the cost lands on recognition accuracy, where nothing shows up until someone thinks to measure it.
Why it matters commercially
On-site voice starts with better input than telephony voice can get, and hardly anybody says so out loud. A visitor speaking into a browser is heard at higher fidelity than a caller on a line, cleaner audio transcribes more accurately, and more accurate transcription is what a good answer is built on.
In voice specifically
Nothing in text works like this. Type a question and the server receives the characters you typed, whatever route they took. Speak one and the network is free to discard part of it on the way, permanently, with no way to reconstruct it at the far end.
Where AsqVox fits
The Orb lives in the browser, so the phone network never enters the picture. Browser capture is usually wideband, and that is a real accuracy edge over a telephony deployment. It is worth saying as an engineering fact and leaving it there.
Visual
What the phone network throws away
A website hears its visitors better than a phone line hears its callers.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
A phone voice channel conventionally spans roughly 300 to 3,400 Hz and is sampled at 8 kHz.
300 to 3,400 Hz at 8 kHzindustry rangeLong-standing PSTN telephony specification, 2026 - It came out of a decision about how little bandwidth intelligible speech needs. Quote it as the published convention it is, never as something anyone metered on a live network.
G.711, the codec the phone network runs on, samples at 8 kHz and 64 kbps, and it is still everywhere.
8 kHz, 64 kbpsindustry rangeITU-T G.711, 2026 - An ITU-T standard, not a vendor number. Its survival is the interesting part: a decades-old codec still dictates what reaches a modern recognizer on a call.
Whisper Large-v3 lands near 2.7 percent word error rate on clean benchmark material and somewhere around 8 to 12 percent on real-world English. Phone audio belongs at the ugly end of that second band.
2.7% clean, 8 to 12% real worldindependentWhisper Large-v3 published benchmark, against reported real-world English performance, 2026 - The clean number is the published model benchmark. The wider band is reported experience rather than one study. The distance between them is what to take away, and bandwidth is one of several things hiding inside it.
Deepgram puts Nova-3 at roughly 5.26 to 6.84 percent word error rate under production conditions.
5.26% to 6.84%vendor claimDeepgram published figures for Nova-3, 2026 - The vendor published it, on conditions the vendor picked. Fine as a sense of what production looks like, useless as a head-to-head against another recognizer.
Nobody has published a trustworthy independent benchmark putting the same models on narrowband phone audio and wideband audio side by side. The penalty is beyond dispute in principle and barely quantified anywhere public.
-no reliable figureSay which way the effect runs and leave it there. A percentage for the narrowband penalty is a number somebody invented, because there is no study to take one from.
Switch-off dates for the old network move around by market and shift again whenever a regulator rules, so there is no global date worth quoting.
-no reliable figureLook up where the relevant country actually stands today instead of recycling a retirement year out of an article nobody dated.
Examples
In practice
A team runs word error rate over its own recordings and finds the phone calls doing far worse than the browser sessions, same recognizer, same speakers. Tracing the media path answers it: crossing a carrier boundary drops the calls to G.711 narrowband. No model swap repairs that. Either boost custom vocabulary so the terms that matter survive, or accept the penalty and build confirmation steps around it.
The everyday version
This is the plain old phone network. Built to carry a voice clearly enough to follow while spending as little bandwidth as it could get away with, it discards anything it can spare. People handle that without noticing. Speech recognition handles it slightly less well, and that is why the same software will sometimes follow a person more easily on a website than on a call.
Usage
Who says it
- Telecoms engineers and regulatory specialists use it exactly.
- Voice AI engineers reach for it mostly to explain why real calls miss benchmark accuracy.
- Buyers hardly ever say it. They say phone line, or landline.
Where it turns up
- Across carrier contracts, regulatory filings, numbering plan paperwork and telephony architecture docs.
- Inside voice AI RFPs it arrives sideways, as questions about phone number support, inbound and outbound calling, and which countries are covered.
Common misuse
- Setting it against VoIP as though they were rivals. Most of this traffic already rides IP. What defines it is the public numbering plan and the interconnect between carriers, not the pipe underneath.
- Planning for wideband on a call. Hand it between carriers and transcoding usually removes that option.
- Comparing telephony and browser agents while ignoring the input quality between them. It is a real variable, and the browser wins it.
Questions people ask
What is the PSTN?
PSTN stands for Public Switched Telephone Network. It is the worldwide arrangement of carriers, national numbering plans, interconnect deals and regulation that lets any phone number dial any other. Internally it is mostly digital and IP now, but the limits it was built with never went away. Reach an ordinary number and you have used it.
Why does a voice agent understand people better on a website than on the phone?
The call simply delivers less signal. A phone channel conventionally covers about 300 to 3,400 Hz at 8 kHz sampling, and several consonant distinctions live outside that window. A browser usually captures wideband, so the recognizer has more to go on. Nothing about the model degraded on the call. Part of what it needed was removed before it got there.
Is the PSTN the same as VoIP?
No, though they are not rivals either. Most of this traffic already travels over IP. What makes it the phone network is the public numbering plan and the interconnect between carriers, not the pipe underneath. A call can be IP from end to end and still be a phone call in every sense that counts.
Can a phone call carry wideband audio?
Sometimes. Certain paths support a wideband codec, and plenty of markets are retiring circuit-switched kit in favor of IP interconnect. Route a call through a few carriers, though, and it usually transcodes down anyway, because the weakest link decides. Plan around narrowband and treat anything better as a bonus.
Last reviewed 30 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.