DTMF
DTMF is the pair of tones a telephone keypad sends — one from a low group, one from a high group, chosen so no human voice produces both at once. It travels three different ways, and a mismatch between the two ends is the usual reason a menu cannot hear your digits. Connect sends DTMF as a live supervisor control; it collects numbers by listening.
Two tones, on purpose#
Each key sounds two sine waves simultaneously: a row frequency and a column frequency. The design predates digital signalling and its cleverness is in the constraint — the frequency pairs were chosen so that speech, music and line noise are very unlikely to produce a valid combination by accident. A single tone would be forged constantly; a pair is nearly never.
That reliability is why the mechanism survived every generation of telephone technology since, and why it remains the only structured input a caller can give a machine without speaking.
Three ways a digit travels#
| Method | How it works | Where it breaks |
|---|---|---|
| In-band audio | The actual tones ride in the audio stream | Low-bitrate compression distorts them; a codec optimised for speech is not obliged to preserve a two-tone chord |
| Named telephone events | The digit is a labelled event in the media stream rather than a sound | Both ends must negotiate the same payload type; a mismatch means digits vanish while audio is fine |
| Out-of-band signalling | The digit is carried in a signalling message | Not universally supported; can arrive out of step with the audio |
The middle method is the usual modern one and the source of the classic symptom: the caller hears their own beeps, the far end registers nothing. The tones the caller hears are generated locally by their handset. Whether anything reached the far end is a separate question that the caller has no way to observe.
Where keypresses still earn their place#
- Navigating somebody else's phone system — an extension, a menu, a conference code. This is the case that never goes away, because the far end is not yours to change.
- Entering digits a caller would rather not say aloud in a public place.
- A fallback when recognition is failing repeatedly: offering keys after two failed attempts is better than a third attempt at the same question.
- Very noisy environments, where a tone survives what speech does not.
What has stopped earning its place is the opening menu. Asking a caller to press one for sales when the system can simply hear 'I want to talk to someone about an order' is a step backwards, and it is the interaction people most often mean when they say they hate phone systems.
Does Connect use DTMF?#
Supported in one specific place, and deliberately not used as the way to collect information from a caller.
Sending. On a realtime call, a colleague watching the live supervisor panel can send digits into the call. Live control rides the guidance log alongside hang-up, transfer and mute, is consumed by the worker's poll, and is never spoken — the same rule that keeps a supervisor's typed guidance out of the caller's ear applies to a control command. That is how a human steers a live call through another organisation's menu without the caller hearing anything about it.
Collecting. Connect listens instead. When a caller says a phone number, an email address or a code, the conversation layer records what it heard as a note naming the value, and then checks the next two lines to confirm the read-back. This is the mechanism that catches a misheard digit — which a keypad would not have had to catch, but which a keypad also cannot collect a spelled-out email address with.
The turn-based carrier path listens with the carrier's own gather element. A production defect worth remembering: when that element was the last thing in the returned document and nothing usable was heard, the carrier ran off the end and hung up. An explicit redirect after it now sends the call back to the turn handler, which answers empty speech with 'sorry, I did not catch that' instead of a dead line.
Questions#
The caller hears the beeps but the system does not react. Why?
Because the beeps are generated by their own handset. Whether the digit reached you depends on the transport both ends negotiated — most often a named-event payload type that does not match. Nothing about what the caller hears tells them, or you, that the digit was delivered.
Should a voice agent offer a keypad option at all?
As a fallback, yes. After two failed attempts at the same piece of information, offering keys respects the caller's time far more than asking a third time. As the primary interface, no — the point of a voice agent is that the caller can say what they want.
Can Connect key its way through another company's menu?
A person supervising a live call can send digits from the panel, so the capability exists under human control. It is not something the model decides to do on its own, and there is no automated menu-navigation feature.