DTMF
In one sentence
DTMF, dual-tone multi-frequency, is the touch-tone scheme that turns every keypad press into its own pair of tones, and it remains the right way to collect the card numbers, PINs and reference codes that speech recognition reliably mangles.
Definition
DTMF is what makes touch-tone dialing work, the beeps a keypad produces. Inside a voice system it is the route a caller takes when they press keys rather than speak.
Nothing has displaced it, because when a value has to be exact, a pressed key wins against a spoken word every time.
Every key sounds two tones at once, one pulled from a low-frequency group and one from a high-frequency group. Sixteen keys, sixteen combinations, each one distinguishable by machine. The scheme is standardized at the ITU-T and has been running since around the middle of the last century.
Why it survives an era of speech recognition
- Accuracy. Press a key and there is nothing to interpret. Say sixteen card digits down a noisy line and there is plenty.
- Privacy. Keying a PIN keeps it out of the room, and out of the recording.
- Compliance. Card handling regularly calls for DTMF capture with suppression, so those digits stay out of the recording and away from any human agent.
- Accessibility, in one narrow but real sense. Callers who cannot speak, or would rather not, still get through.
The two transmission modes, and why the difference bites
- Inband puts the tones in the media stream as ordinary audio. Compression can smear them past the point of detection. G.711 copes. Squeeze the bitrate hard enough and it stops coping.
- Out-of-band sends each press as a discrete event instead, normally as an RFC 2833 or RFC 4733 telephone-event payload. Far steadier, and the approach to reach for.
- Detect inband on a compressed path and the failures arrive at random, which makes for one of the more maddening bug classes in telephony.
DTMF in AI voice agents
- Plenty of agents accept both on one field. Speech leads, the keypad is offered for numbers, and repeated misrecognition pushes the caller across to keys.
- The pairing genuinely outperforms either half on its own. Talk for the open-ended part, press for anything that has to be exact.
- No keypad exists on a website, so a text field takes the role. It covers the same three needs: exactness, privacy, and a way through for anyone who cannot speak.
Common misconception
That speech recognition retired it. Card numbers, PINs and reference codes are still its territory, and payment rules often insist on it outright. Strip it out to look modern and you lose accuracy and compliance standing in the same move.
Why it matters commercially
It is a minor line item that stops deployments dead when it is missing. Put an agent anywhere near payments, account numbers or identity checks and somebody will ask. In regulated sectors that answer settles whether the project happens at all.
In voice specifically
A sixteen-digit string is the wrong thing to ask anyone to say. It is slow, it invites mishearing, and it broadcasts to whoever happens to be nearby. Sooner or later every voice interface needs a route that is not speech for exact and private values, whether that turns out to be a keypad or a text field.
Where AsqVox fits
A browser has no keypad, so DTMF has nothing to attach to. The Orb's text chat fallback fills the same role, handing a visitor the exactness and privacy a caller gets from a keypad. Worth drawing the comparison, because it recasts text fallback as a real capability instead of a consolation prize.
Visual
When keys beat speech
On a website there is no keypad. Text entry does the same job.
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.
Every key sounds one low-group tone and one high-group tone together, which yields sixteen combinations a machine can separate.
Sixteen key combinationsindustry rangeITU-T standardized DTMF signaling scheme, 2026 - A standard, not a benchmark. Pairing the tones is what lets a detector pick a keypress out of speech and line noise instead of guessing.
On SIP paths, out-of-band presses normally ride as telephone-event payloads, defined in IETF RFC 2833 and in RFC 4733 after it.
RFC 2833 and RFC 4733industry rangeIETF RFC 2833 and RFC 4733, 2026 - Write the RFC number into the RFP. It separates a vendor that signals presses as events from one quietly hoping the tones outlive the codec.
Capture with suppression is a routine control in card handling, gathering digits while keeping them out of recordings and away from human agents.
-industry rangePCI DSS obligations, rather than the DTMF standards themselves, 2026 - Payment card compliance imposes this, not anything written into the DTMF specification. Attribute it correctly when a security questionnaire asks.
Nobody has published a trustworthy figure comparing keypad entry against speech for numeric fields in production, even though it would be straightforward to measure.
-no reliable figurePractitioners come down heavily on the keypad for exact strings. Report that as consensus, keep it away from decimal places, and refuse the vendor deck that turns it into a percentage.
Examples
In practice
An agent taking card details over speech recognition posts a dismal completion rate on the card number step, because sixteen digits down a compressed line will usually go wrong somewhere. Move that one step to out-of-band capture with suppression and completion climbs sharply, while the requirement that card digits stay out of the recording is satisfied in the same change.
The everyday version
This is pressing buttons instead of talking. Reading out a sixteen-digit card number is slow, easy to get wrong, and uncomfortable with anyone in earshot. Typing it is quick, precise and private. Well-built systems let people talk through the easy parts and type the fiddly ones.
Usage
Who says it
- Telephony engineers and IVR designers say it as a matter of routine.
- Payment compliance specialists say it, almost always about suppression.
- Buyers seldom do. They say keypad entry, or touch-tone.
Where it turns up
- Alongside payment handling, PCI DSS scope, recording and suppression, and accessibility requirements.
- In regulated sectors it lands as a hard requirement, not a nice-to-have. A vendor without out-of-band capture and suppression is usually out of the running immediately.
Common misuse
- Settling for inband detection. Compression makes it flaky, and the flakiness is intermittent and awful to reproduce.
- Cutting keypad entry because recognition improved. For exact strings and private values it is still the better tool.
- Reading its absence in a browser as a hole in the product. A text field covers the identical ground.
Questions people ask
What does DTMF stand for?
Dual-tone multi-frequency. Press a key and it sounds two tones at once, one from a low-frequency group and one from a high-frequency group, giving sixteen combinations a machine can separate cleanly. The ITU-T standardized it, and it has been in service since roughly the middle of the last century.
Do AI voice agents still need DTMF?
Wherever a value has to be exact, yes. Card numbers, PINs and reference codes are the usual cases, and card handling regularly demands capture with suppression so those digits stay out of recordings and away from human agents. Many agents now take speech and keys on the same field, moving to keys once recognition has failed a few times.
What is the difference between inband and out-of-band DTMF?
Inband carries the tones as audio in the media stream, so compression can smear them past detection. G.711 copes; squeeze the bitrate hard and it stops coping. Out-of-band, meaning RFC 2833 or the RFC 4733 that replaced it, sends each press as a discrete telephone event. Steadier, and the approach to prefer.
How does a website voice agent collect a card number without a keypad?
Through a text field. No browser has a keypad, so the text input takes that role and covers the same three needs a keypad covers on a line: exactness, privacy, and a way through for anyone who cannot or would rather not speak. Calling its absence a gap mistakes what the mechanism was ever there to do.
Last reviewed 30 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.