Detecting voicemail in under two seconds: what exists today

Why it is hard

When a person picks up, they say a short word — “Hello?” — then stop and wait. A voicemail does the opposite: it talks for a long time, expects no answer, then beeps. On paper the difference is clear. On the phone it is not: the greeting sometimes starts with silence, sometimes with a carrier sentence, sometimes with the person's own voice (“Hi, you've reached…”). The sound is narrow (telephone band, 8 kHz), the first instants are often clipped, and the agent must decide quickly: talking to a voicemail as if it were a person wastes time and money; hanging up on a person mistaken for a voicemail is worse.

Time counts twice. On our September 2026 calls, half of the calls answered by a human last ten seconds or less. A decision that comes at eight seconds therefore comes, for many calls, after the conversation has already happened.

The three ways to recognise a voicemail

1. Wait for the beep

The oldest method listens to the signal for the tone that ends the greeting. It is reliable when the beep is clean, but it can say nothing before the beep arrives — that is, after the whole greeting. It fails when the beep is faint, missing, or buried in the compressed sound of the line; Plura details these failure cases in an article cited below.

2. Listen to the cadence

The second method listens not to words but to rhythm: how long the other party talks, how long they stay silent, whether they wait for an answer. A person says a few words then leaves a gap; a voicemail keeps going. This is the principle of the detectors built into telephony carriers, and of the paper published on arXiv in April 2026 (“Real-Time Voicemail Detection in Telephony Audio Using Temporal Speech Activity Features”), which shows that a real-time decision can be made from speech-activity cues alone. The method needs a few seconds of listening to be sure: too little, and a talkative person looks like a voicemail.

3. Transcribe and classify

The third method transcribes the first words and classifies them: “you've reached the voicemail of” leaves no doubt, and neither does “Hello, yes?”. It is the most precise when the transcription is good, but it depends on the quality of speech recognition over the phone, and it has to wait for words. Recent voice-agent vendors generally combine cadence and transcription, keeping the beep as a safety net.

What the vendors publish

Twilio, whose documentation is the most complete, states that its detector returns a result in about four seconds with default settings, and that its maximum delay can be set between 3 and 59 seconds (30 by default). Its FAQ announces accuracy “close to 100%” in the United States and acknowledges it is lower internationally; no per-language figure is published. Regal.ai describes an in-house model it presents as better than the market, measured on its own North American calls. Vapi documents several strategies its customers can choose from. None of these pages gives a measurement on European calls, nor deals with the voicemail greetings of French or Spanish carriers.

Where ASDIC stands

We do not sell a detector; we are building a model that must decide earlier, and we publish our measurements as they are. In September 2026, on Nodical's production database (60,287 calls answered by a human, 14 campaigns, without reading the content of conversations), the voicemail is recognised at 8 seconds (median), and at 14 seconds for 9 calls out of 10. That is too late, and it is exactly what the 2027-2028 programme must change.

Our target is written as a target: 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 detail of the September measurement, its method and its limits are in the first notebook entry.

Why Europe is harder than the United States

Three reasons, which add up. Voicemail greetings are more varied: each carrier has its own, in each language, and many people record theirs. Calls are multilingual: one model has to hold in French, Spanish and English without building one per language. Finally, the published measurements were made elsewhere: a detector “close to 100%” on American voicemails has never heard a French carrier greeting. That is why we measure at home, and why we will publish the method with the results.

What it changes for a calling campaign

Every second spent talking to a voicemail is billed, and every person hung up on by mistake is a lost contact. For a team that lets a voice agent make its calls, “human or voicemail” is the first decision of the call, and it conditions all the others: when to speak, what to say, when to hand over. Nodical explains on its site what it changes for a calling campaign.

And in practice? what it changes for a calling campaign

Sources

accessed on 1 October 2026.

1. Twilio, « Answering Machine Detection » (documentation) — https://www.twilio.com/docs/voice/answering-machine-detection
2. Twilio, « AMD FAQ & Best Practices » — https://www.twilio.com/docs/voice/answering-machine-detection-faq-best-practices
3. arXiv 2604.09675, « Real-Time Voicemail Detection in Telephony Audio Using Temporal Speech Activity Features » (avril 2026) — https://arxiv.org/html/2604.09675v1
4. Plura, « Voicemail beep detection: how it works and why it fails » — https://www.plura.ai/articles/voicemail-beep-detection
5. Regal.ai, « How we built a voicemail detection model » — https://www.regal.ai/blog/how-we-built-an-answering-machine-detection-model-that-outperforms-the-industry
6. Vapi, « Voicemail detection » (documentation) — https://docs.vapi.ai/calls/voicemail-detection

Corrections

No correction so far.