Answering Machine Detection for Arabic Dialects

Answering machine detection is the least glamorous control on an outbound floor and the one that quietly decides how many conversations your agents get. Every wrong verdict costs something. Call a machine a human and an agent listens to a recording for twenty seconds. Call a human a machine and you hang up on a customer who answered. In Arabic-speaking markets the traditional approach gets both wrong more often than teams expect.

Why Beep Detection Fails on Arabic Greetings

Classical AMD is an audio-timing technique. It waits for the far end to speak, measures how long the speech runs, waits for silence, and then listens for a tone. If the tone arrives, it declares a machine. That chain of assumptions holds reasonably well on a market where voicemail is standardised and greetings are short. It holds badly across the Gulf and the Levant, for four separate reasons.

• Many operator voicemail systems in the region produce no clean beep at all, or produce one after a carrier announcement that the detector has already given up on. A detector waiting for a tone that never comes has no verdict to give.
• A live Arabic greeting is frequently long. A caller who answers with a full greeting, a welcome and a question has spoken for several seconds before the first natural pause, which is exactly the pattern a timing detector reads as a recorded message.
• Recorded greetings in the region are often very short, sometimes a name and nothing else. Short speech followed by silence is the pattern a timing detector reads as a live person.
• Mobile networks in the region frequently insert their own announcements, in Arabic and then in English, before the subscriber's own greeting. The detector measures the announcement rather than the greeting.

The common thread is that timing carries no information about who is speaking. The words do. A detector that cannot read them is guessing, and it guesses worse in Arabic than in the market it was tuned on.

What Transcript Classification Does Instead

DialerBee's AI AMD does not wait for a tone. The moment the far end picks up, the opening audio is transcribed in real time, and a classifier scores the resulting text rather than the waveform. The difference is that the text is unambiguous in a way the timing never was. A person saying "aywa, meen maai?" and a recording saying "the subscriber you have called is not available, please leave a message after the tone" are not close to each other as text, however similar their timing profile.

Four steps run on every answered call.

• Transcribe. The opening seconds of audio are converted to text as they arrive.
• Detect the language and dialect. This is automatic, so a mixed-language list does not need to be split into a campaign per language before it can be dialed.
• Classify. The classifier returns a human or machine verdict together with a confidence score.
• Act. The campaign threshold decides what that confidence score means for this campaign, so a collections list and a customer-satisfaction list can sit at different levels of caution on the same platform.

The verdict is produced from the opening moments of the call, while the greeting is still playing, so the decision to connect or drop happens inside the pickup phase rather than after it. We do not publish a single latency figure, because the real number moves with the carrier, the language and how quickly the greeting starts. Ask for measurements taken on traffic that resembles yours.

Dialect Coverage Across the Region

Arabic is not one target. A classifier trained on formal broadcast Arabic will read a Gulf voicemail greeting and a Levantine one very differently, and neither resembles the Egyptian or Maghrebi equivalent. DialerBee treats the dialect as part of the language rather than as noise around it, and the Arabic coverage spans Gulf, Levantine, Egyptian and Maghrebi speech.

In practice that matters most at the boundaries. A Jordanian mobile greeting and a Saudi mobile greeting use different verbs for the same idea. A Gulf caller answering a business call often opens with a formal religious greeting that a narrow model treats as scripted, and therefore recorded. A Maghrebi greeting may carry French vocabulary inside an Arabic sentence. All four are ordinary traffic for a regional BPO, and all four are places where a single-dialect model loses conversations.

Arabic sits inside the platform's 11 supported languages, so a floor that dials Arabic, English, French and Urdu in one week runs one detection stack rather than one per market.

Thresholds, Not Absolutes

A confidence score is only useful if you decide what to do with it, and the right decision differs by campaign. On a late-stage recovery campaign, dropping a live debtor is expensive and leaving a voicemail is cheap, so you set the threshold to be cautious about declaring a machine. On a high-volume appointment reminder campaign the balance reverses.

Sensitivity thresholds are therefore configured per tenant and per campaign rather than platform-wide, which is what makes the control usable for a BPO running several clients with different tolerances. AI features can be controlled by feature flags, so AMD can also be enabled or disabled per campaign, per tenant or globally, and the platform runs at full capability with it switched off.

Two more configuration facts are worth knowing before an information-security review. AMD classification processes a short initial segment of call audio, and raw audio used for classification is not stored long term. What is retained is the classification result, the confidence score and the model version.

The Agent Feedback Loop

No classifier is right every time, and the people who find out first are the agents. When the model gets it wrong in either direction, the agent flags it with a single click from the desktop, during the call, with no form and no ticket. Those corrections feed tenant-scoped tuning workflows, which means the signal improves detection for your carriers, your regions and your campaign types and is not shared across tenants.

This works only if supervisors treat the override as part of the job rather than as an exception report. Two habits help. Brief agents that flagging a wrong verdict takes one click and is expected, not a complaint about the system. And read the per-campaign, per-carrier and per-language accuracy breakdowns weekly, because a carrier that has changed its voicemail announcement shows up there as a pattern long before anyone reports it as a problem.

Beyond Arabic: The Other Languages on the Same List

Language-aware detection is not translation. It is recognising how people and machines actually speak in a market, and the variation that makes Arabic hard exists in every other language on a regional list. Spanish splits between Latin American and European speech. French carries European, Canadian and North African patterns. German runs formal and informal registers plus Austrian and Swiss variants. Turkish has its own regional patterns. English alone arrives in US, UK, Indian, Philippine and Gulf-accented forms, which is most of a BPO's agent-facing traffic before a single Arabic number is dialed.

This is why the detection stack matters more than any single language claim. A floor that dials Arabic on Sunday, French on Monday and Urdu on Tuesday should not be running three detection configurations and reconciling three sets of accuracy numbers. All 11 supported languages share one classifier and one reporting view, and dialect is treated as part of the language in each of them rather than as noise.

How to Measure Detection Properly

A single headline accuracy figure is close to useless, because it averages away the place where your losses are concentrated. Measure this instead, and insist on the same breakdown from any vendor you are evaluating.

• False positives and false negatives separately, and per carrier rather than per region. A 90 percent regional average can hide 65 percent accuracy on the operator carrying most of your volume.
• Agent override rate, which tells you what the floor thinks of the verdicts, and how fast tuning has something to work with.
• Detection time, alongside the connect decision, since a correct verdict that arrives after the customer has hung up is not a win.
• Results split by language and by campaign type, because a recovery list and a reminder list will not behave alike even on the same numbers.

Two habits protect you from buying a number rather than a capability. Test on your own traffic, through your own trunks, to your own destination carriers, since accuracy on demo or synthetic audio rarely survives contact with production. And pilot by campaign type instead of flipping one global setting, reviewing false positives first because a live person who was dropped is the expensive error and the one that never appears in an agent's queue.

When a route, a list source, a region or a script changes, run the review again. Carrier voicemail behaviour is not a fixed target, and a detection setting that was correct last quarter is only a hypothesis this quarter.

How AMD Fits the Rest of the Floor

AMD is not a standalone product, and its value comes from what happens after the verdict. When a machine is detected the agent can drop a prepared voicemail message with one click and move to the next call while the message is still playing, using templates generated by text to speech in a voice you chose and synced to every media node so a drop works wherever the call is running. Alternatively the contact can be re-queued automatically under the campaign's retry rules.

The verdicts also feed the pacing model. When AMD filters voicemail out of the connected calls, the predictive engine adjusts its estimate of the true live-connect rate, so pacing reflects the conversations agents will actually take. AMD works with any carrier connected over SIP, which matters in a region where most operators bring their own trunks.

Frequently Asked Questions

Why does beep detection struggle with Arabic voicemail?

Because it reads timing rather than words. Many operator voicemail systems in the region play no usable beep, live Arabic greetings are often long enough to look like a recording, recorded greetings are often short enough to look like a person, and network announcements in Arabic and English arrive before the subscriber greeting. None of those problems is solvable by better tone detection.

How does transcript-based AMD classify a call?

The opening seconds of audio are transcribed in real time, the language and dialect are identified automatically, and a classifier scores the text and returns a human or machine verdict with a confidence score. The campaign threshold then decides what to do with that score, so different campaigns can act on the same verdict differently.

Which Arabic dialects are covered?

Coverage spans Gulf, Levantine, Egyptian and Maghrebi speech, and Arabic is one of the platform's 11 supported languages. Dialect is treated as part of the language rather than as noise, which matters most on mixed regional lists where a Jordanian, a Saudi and a Moroccan greeting all appear in the same campaign.

Can I set different AMD sensitivity per campaign?

Yes. Sensitivity thresholds are configured per tenant and per campaign, so a late-stage recovery campaign can be cautious about declaring a machine while a high-volume reminder campaign is not. AI features can be controlled by feature flags, so AMD can also be turned off per campaign, per tenant or globally.

What happens when an agent disagrees with the verdict?

The agent flags it with one click from the desktop during the call, with no form and no ticket. Corrections feed tenant-scoped tuning workflows, so the signal applies to your carriers, your regions and your campaign types. Supervisors should also review the per-campaign, per-carrier and per-language accuracy breakdowns regularly.

Is call audio kept for AMD processing?

AMD classification processes a short initial segment of call audio, and the raw audio used for classification is not stored long term. What is retained is the classification result, the confidence score and the model version, which is what an information-security review normally needs to see.