The voicemail beep: why it never comes when you expect it

Two decisions, not one

When a voice agent reaches a voicemail, it has to settle two things. First: person or machine? That is voicemail detection, which has its own page. Then, if it wants to leave a message: at what moment does the voicemail start recording? The first decision can be made from the opening words of the greeting. The second only at the end, and that end is usually marked by a beep. That is the whole problem: the beep never comes at the same moment, never sounds quite the same, and sometimes never comes at all.

Plura, which sells outbound dialling software, sums up the difference: voicemail detection says who picked up, beep detection says when recording starts.

What you hear before the beep

• A carrier greeting, along the lines of “The person you are calling is not available. Please leave a message after the tone.” The same for millions of subscribers, but different from one carrier and one country to the next.
• A personal greeting: three seconds or forty, with music, a pause in the middle, sometimes a “Hello?” that mimics a real answer.
• Both in a row: the personal greeting, then the carrier's sentence.
• A network message (phone switched off, out of coverage): neither a person nor a voicemail, so no beep, since nothing is recording.

Twilio gives an order of magnitude for short greetings: typically 1 to 2 seconds of speech, then 1.2 to 2.4 seconds of silence, then more audio before the beep. A detector that takes that silence for the end of the greeting gets it wrong; Twilio acknowledges that its own may take a two-second greeting for a person picking up. At the other extreme, its FAQ advises leaving the beep time to arrive, because some greetings are very long.

A beep with no standard

Twilio says so plainly: what counts as a voicemail “beep” varies greatly from one destination to another, and it sometimes overlaps with ordinary network tones to the point of being indistinguishable from them. Plura agrees: beep frequency and duration are not standardised, a detector tuned on one carrier's voicemail often misses another's, and many modern voicemail systems emit no beep at all — they use silence, a spoken prompt, or both.

The line adds its own flaws: narrow band, compression, lost packets (see the page on 8 kHz audio). A beep can arrive clipped; music inside the greeting can sound like a beep.

What Twilio's DetectMessageEnd option does

Twilio offers two modes. Enable answers as soon as it knows who picked up; it is mostly used to pass only human calls to an agent. DetectMessageEnd is made for leaving a message: it answers immediately if it hears a person, but with a machine it waits for the end of the greeting. It then returns one of three results: greeting ended with a beep, with silence, or with something else (a tone, an error message). That third result is also the fallback when the maximum delay runs out before the greeting ends. Unless you need to tell them apart, Twilio writes, all three mean the same thing: a machine answered.

That maximum delay can be set from 3 to 59 seconds (30 by default); too short, and listening stops before the beep. And it is for this mode, with default settings, that Twilio announces accuracy “close to 100%” for US destinations, lower internationally because voicemail tones differ from one country to another. No figure is published for the United Kingdom, France or Spain.

Why waiting for the beep is expensive

Waiting for the beep means listening to the whole greeting: far too late to decide “person or machine”. The paper published on arXiv in April 2026 on real-time voicemail detection points the same way: adding beep-based features did not improve its best configuration and substantially increased latency. The beep does not help you know who you are talking to; it remains the only cue for knowing when to speak. On our September 2026 calls, voicemail is recognised at 8 seconds (median), and at 14 seconds for 9 calls out of 10. We do not measure the moment of the beep, so we do not know how much time separates, on our calls, recognising the voicemail from the opening of the recording.

Leaving a message: what it implies

Started too early, the message gets clipped: it begins during the greeting, and the voicemail only records the end, or silence. Plura advises waiting for the end of the beep plus a short margin, especially when the greeting ended in silence. Started too late, it opens with a blank that sounds like an empty call. Hence a few simple rules:

• say what matters first — who is calling, why, how to call back — so that a clipped start does not lose the meaning;
• keep it short;
• leave nothing when the end of the greeting is uncertain, rather than a half-recorded message;
• ask whether voice is the right channel: a text arrives whole and keeps the callback number.

Where ASDIC stands

Our model must first answer the first question earlier: the target is to recognise a voicemail in under 2 seconds, with less than 3% error. It is not reached; it will be published with the measurement that proves it, not before. The second question, the moment recording starts, needs hand-annotated calls marking, for each voicemail, the moment of the beep or its absence. Until that work is done, we publish no figure about the beep. The September measurement and its limits are in the first notebook entry.

When writing works better

A voice message depends on a beep nobody controls; a written message does not. Nodical explains on its site, about missed calls, when it is better to send a text, and what a text does not replace.

And in practice? when it is better to send a text

Sources

accessed on 1 October 2026.

1. Twilio, « Answering Machine Detection » (documentation : modes Enable et DetectMessageEnd, réglages, annonces courtes) — https://www.twilio.com/docs/voice/answering-machine-detection
2. Twilio, « AMD FAQ & Best Practices » (résultats machine_end_*, définition variable du bip, précision de DetectMessageEnd) — https://www.twilio.com/docs/voice/answering-machine-detection-faq-best-practices
3. Plura, « Voicemail Beep Detection: AMD vs. Beep Timing Explained » (17 août 2026, mis à jour le 29 septembre 2026) — https://www.plura.ai/articles/voicemail-beep-detection
4. arXiv 2604.09675, « Real-Time Voicemail Detection in Telephony Audio Using Temporal Speech Activity Features » (avril 2026) — https://arxiv.org/abs/2604.09675
5. ASDIC, carnet n° 1 : « Septembre 2026 : première mesure » — https://asdic.ai/carnet/2026-09-premiere-mesure

Corrections

No correction so far.