PSTN
The PSTN is the regulated, interconnected telephone network that ordinary numbers live on. Software cannot join it — you buy access from a licensed carrier, per country, under that country's rules. What it gives you is reach to anyone with a phone. What it charges is per minute, narrowband audio, and a signalling channel with almost nothing in it.
What you are actually connecting to#
Most of the network has been packet-switched internally for years, but the properties that matter to a software system come from its origins as a circuit-switched, nationally regulated utility. Numbers are allocated under national plans and expressed internationally in E.164 form — a country code, a national number, at most fifteen digits. Interconnection between operators is commercial and regulated. Portability means a number's prefix no longer reliably identifies who carries it.
The practical consequence is that there is no global telephone API, and there will not be one: eligibility for a number, the documentation you must supply to get one, what you may present as calling line identity, and what you may do with an automated call all vary by country. A design that assumes otherwise fails at the second country.
The constraints a call imposes on software#
- Narrowband audio
- The classic voice path is 8 kHz sampled, and consumer-grade at that. A speech model on a phone call is working with materially less signal than the same model on a laptop microphone, which is why phone-call accuracy is never the accuracy in a demo.
- Per-minute cost
- Charged by destination and often asymmetric. Cost belongs on the call record, not in a monthly guess.
- No rich data channel
- A caller can send you digits and speech. There is no structured payload, no attachment, no read receipt.
- Call state is hearsay
- You know what the carrier tells you, when it tells you. A call your system believes is active may have ended two minutes ago.
- Latency is cumulative
- Network transit, jitter buffering and any transcoding all add to the gap before your agent can even begin to think.
Does Connect use the PSTN?#
Yes, on every real call. A carrier supplies the PSTN leg — Vobiz on the live account — and inbound and outbound calling both run in production. Connect adds a realtime path on top, where the number is linked at the carrier to a SIP trunk and a speech-to-speech model holds the conversation, but the leg that reaches a human's handset is still the public network.
The provider is deliberately confined. One adapter module knows a carrier's signature scheme, its XML dialect and its webhook shape, and it is the only place a provider's own words appear. The core knows what a line is, what an event means and who may be called when; providers differ there only through a capability table, asked with telephony.can(...). Everything above names no provider at all, and a test asserts that rather than trusting it.
Three things the network can do that this account does not: call recording is a provider capability that is not enabled here; a completed warm handover to a colleague depends on a provider capability that is not enabled either, so an escalation queues rather than completes; and the live carrier carries no SMS at all, which the Phone screen states rather than presenting a message thread that cannot send.
What this means for expectations#
A voice agent is not a chat window with audio attached. The floor on the realtime path is the model's first token plus its end-of-turn detection — 3.3 s median reply on the best measured call, not the one to two seconds a demo video implies. That floor is a property of the model and the network together, and no setting removes it.
It also means every design question about a phone call is a question about time. What can be said while the caller is still speaking, what must wait, what is worth the round trip — these matter here in a way they never do on email, where a two-second delay is invisible.
Questions#
Can software place a call without a carrier?
Not to a telephone number. You can carry audio between two internet endpoints with no operator involved, but the moment one end is an ordinary phone, a licensed operator terminates that call and charges for it. Choosing a carrier is therefore a product decision, not an implementation detail.
Why does a phone call sound worse than a video meeting?
Narrowband audio. A conference tool can negotiate wideband between two computers; a call to a handset is constrained by the oldest link in the path. This affects recognition as much as it affects how a voice sounds — quiet speech, background noise and unfamiliar names all cost more accuracy here.
Does the number I dial from matter?
Yes, in two ways. Presenting a calling line identity you do not control is prohibited in most jurisdictions, and separately, recipients answer numbers they recognise. Both point at the same practice: call from a number the business genuinely holds and publishes.