Skip to main content
Callers can type on their phone keypad instead of speaking. The digits arrive as a normal caller turn, so your agent handles them with the same prompt it uses for speech - no extra configuration, no separate flow. Reading numbers aloud is where speech recognition fails worst. A mis-heard account number sends the caller round the loop, and a stray word can cut the agent off mid-sentence. A keypad press is exact.
Keypad input is off by default. Turn it on per agent in the dashboard, or per number through the API.

​
Turning it on

​
Per agent (dashboard)

Open the agent, find Call actions, and switch on Keypad input. Every number on that agent accepts keypad entry from the next call onwards.

​
Per number (API)

A single number can override its agent - useful when one line needs keypad entry and the rest do not.
null is not the same as false. null means “no opinion, follow the agent”, so a null number under an agent with keypad input on will accept keypad entry. To switch one number off while the agent stays on, send false, not null.
See Update phone number.

​
What your agent receives

Digits are handed to the model as a caller turn, tagged so it knows they were typed rather than spoken:
That tag matters. The model should read a typed account number back to confirm it, and it should not treat a lone 1 as the caller saying the word “one” mid-sentence.
Ask for keypad entry in your prompt and the model will use it naturally: “If the caller needs to give a reference number, ask them to type it on the keypad and then read it back to confirm.”

​
When an entry is finished

Two things end an entry: The three-second timer restarts on every press, so a caller slowly reading a long card number is never cut in half. There is no maximum length.
The three-second pause and the # send key are fixed and not configurable today.
* is accepted as an ordinary key and comes through as part of the value.

​
Where it works

​
Behaviour you can rely on

  • The agent never waits for speech to process digits. Once an entry is finished it goes straight to the model - a caller who types ten digits and says nothing still gets a reply.
  • Typing does not trigger the silence hang-up. The inactivity timer resets on every entry, so a slow typer is not cut off.
  • A half-typed entry is discarded when the call ends. Digits can never arrive after hangup.
  • Junk is ignored, never fatal. Anything that is not a single 0-9, * or # is dropped rather than raising an error, so a malformed frame cannot take down a live call.

​
Common uses

  • Account or reference numbers - exact where speech is not
  • Menu choices - “press 1 for sales, 2 for support”
  • Postcodes and card numbers - long strings that speech recognition mangles
  • Dates of birth - digits with no spelling ambiguity

​
Testing it

1

Turn it on

Switch on Keypad input for the agent, or PATCH dtmf_enabled: true on the number.
2

Call the number

Ask the agent something, then type a few digits on your keypad.
3

Press hash

The entry is sent the moment you press #. Without #, it sends about three seconds after your last press.
4

Check the transcript

The call transcript shows the entry as a caller turn: [Caller pressed on keypad: 5869].