How Parallel Dialers Detect Voicemail vs a Live Person: The Real Numbers Behind Answering Machine Detection

By Renata Osei · Reviews Lead · November 12, 2026 · 105 sources

Parallel dialers classify every answered call with software called Answering Machine Detection (AMD), which listens to the first seconds of audio after a call connects and returns a verdict: human, machine, or unknown. The dialer then either bridges the human to a rep or drops and leaves a message for the voicemail. Detection is always a trade between speed and accuracy, and every vendor, including CallCloud, Nooks, Orum, Kixie, and Salesfinity, sits at a different point on that curve.

The honest answer about how this works involves timing thresholds, acoustic models, and marketing claims that are rarely comparable. This article covers what AMD actually measures, why it takes the time it takes, which published accuracy numbers you can trust (mostly none, unqualified), and how to test detection before you sign anything.

Parallel dialers use Answering Machine Detection, software that analyzes the first seconds of audio after a call connects and classifies it as human, machine, or unknown. Most detectors do not understand the words being spoken. They look for patterns in silence, greeting length, pauses, speech bursts, tones, and acoustic signatures that statistically separate people from recordings.

Once the verdict lands, the dialer acts on it. A live person is bridged to the rep. CallCloud, for example, keeps the rep on a live audio bridge for the whole session, so the connect step after a human verdict is nearly instant rather than a fresh transfer. A machine verdict means the call is either dropped or handed a prerecorded message.

Four generations of method are in active use:

1. Timing heuristics, built on silence and greeting length.
2. Tone and audio fingerprint detection.
3. Machine learning acoustic models.
4. Speech-to-text semantic classifiers that recognize phrases like "leave a message after the beep."

The single most useful thing to understand before comparing any two dialers: detection is a speed-versus-accuracy trade, and every vendor picks a different operating point on that curve. A detector that decides in half a second will make more mistakes than one that listens for five seconds. Neither number is wrong. They are different bets.

The detector listens for a small set of measurable signals, and the vendor's published configurations make the mechanics surprisingly concrete.

• ·Initial silence. A long gap before any greeting suggests a machine. Asterisk's sample configuration flags a call as machine after 2,500 ms of initial silence.
• ·Greeting length. An unbroken, long greeting suggests a recording; a short one suggests a human. Genesys' forwarding FAQ defines the live-person pattern as under 2,200 ms of speech followed by at least 700 ms of silence.
• ·Pause and end-of-greeting analysis. The detector waits for enough post-speech silence to conclude the greeting is over. Twilio's speech-end threshold defaults to 1,200 ms.
• ·Speech burst counting. Asterisk counts voiced segments separated by silence, with a sample maximum of 2 words and a 5,000 ms maximum word length before declaring machine.
• ·Tone and beep detection. FreeSWITCH's mod_avmd watches for a stable single frequency roughly between 440 and 2,000 Hz. This is excellent for placing a voicemail message but usually too late to protect a fast human connection.
• ·Audio fingerprints. Matching incoming audio against known carrier and voicemail recordings, as Genesys lists among its three mechanisms. New or personalized greetings can miss the library.
• ·Screening-aware state machines. With call screening spreading, including Apple's iOS screening, detectors like Telnyx's screening mode restart AMD after a screening event so a real person behind the screener is not written off as a machine.
• ·Human classification. Some products skip automatic classification entirely. PhoneBurner has the rep click a voicemail button once they hear a mailbox; ConnectAndSell and MonsterConnect use humans to listen and transfer. It costs labor, but it avoids model errors.

Traditional AMD is slow by design: it must hear enough audio to classify. Twilio reports roughly 4 seconds on average with default settings; Plivo publishes 3 to 5 seconds; Readymode offers a roughly 3-second Quick check versus a 7-second Rigorous check.

Twilio's four tuning controls expose the trade directly:

Here is the dial in concrete form: raising the speech-end silence threshold from 1,200 ms to 2,500 ms adds about 1,300 ms to every human detection, in exchange for catching more tricky voicemails. Every threshold is a purchase of accuracy with seconds.

Modern streaming models decide earlier. Nooks publishes predictions every 100 ms and says most calls resolve in under one second. LiveKit publishes an 840 ms median time to verdict with 94.7% accuracy in its multiclass benchmark.

The trap when comparing speed claims is that six different clocks get marketed as "latency." CallCloud's 22 ms figure is answer-to-rep bridge time on a persistent live connection, not an AMD verdict time. Nooks' 100 ms is prediction cadence. JustCall's 4 to 5 seconds includes analysis plus connection. Retell's under 30 ms is incremental processing overhead. Ask each vendor which clock they are quoting, and demand a call-level median measured from answer to verdict.

That gives you the decision rule for this whole category: separate connection speed from detection speed. A dialer can connect a confirmed human instantly while its classifier is average, or classify brilliantly while taking seconds to hand the call over. You have to evaluate both numbers, and a vendor who publishes one and not the other has told you something.

Because the claims are not on the same scale. Convoso advertises "up to 97% accuracy." MightyCall claims 97% precision within 3 seconds. Voximplant publishes 99.26%. Plivo's own guidance says 60 to 75%. JustCall says typically 75% or higher. Twilio's near-100% statement applies only to DetectMessageEnd on US destinations with default settings.

Precision, accuracy, recall, and macro-F1 measure different things, and vendors do not all use the same positive class. A 99% figure with no stated false-machine rate tells you almost nothing.

Field results can collapse. VM Hunter's 4,000-call, one-week, human-labelled field test found 61% of machines were caught by neither detector it tested. Its acoustic comparator's held-out 0.977 macro-F1 degraded badly in the field, passing 78.7% of machines at one threshold while losing 4.6% of live people.

The hardest case is the short greeting. A two-second voicemail greeting can pass as human. A person who says "hello," pauses 2 seconds, and repeats "hello" can be classified as a machine by timing-based detectors.

AMDY illustrates why you should read claim pages carefully: its pricing page advertises 99% accuracy while its own features page states it lacks a labelled ground-truth set and cannot honestly quote one.

Connect rates are not detection accuracy either. Salesfinity's published 10.1% connect rate across 1,315,034 dials is a workflow outcome shaped by list quality and pacing, not a classifier benchmark.

Two errors, two different costs. A false machine verdict, meaning a human classified as voicemail, silently discards a prospect and can count as an abandoned or silent call. A false human verdict wastes rep time talking over a recording.

US abandonment rules make the false-machine error a legal issue, not just a sales one. 47 CFR 64.1200 and the FTC TSR safe harbor require connecting a live rep within 2 seconds of the completed greeting, with no more than 3% of live answers abandoned over rolling 30-day periods. Ofcom's UK policy treats false-machine hangups caused by AMD as potential silent-call misuse, and expects records kept for at least 6 months.

The engineering consequence: if a false-machine error drops a human before a rep is connected, the legal clock and the business cost both fire. Faster human detection reduces dead air, but aggressive machine rejection can silently throw away people and quietly push a campaign over the abandonment ceiling.

For tuning, the split is which error you can afford. When prospects are valuable, bias toward passing uncertain calls to the rep or listening longer. When volume is cheap and voicemail prevalence is high, a tighter machine threshold saves rep time. The right setting depends on your economics, not on any universal accuracy target.

PickNo. 01

CallCloud

$125/seat/month (Parallel), $60/seat/month (Power)

CallCloud separates the bridge from the classifier: the rep stays on a live audio bridge, so once a human verdict lands the connect step is nearly instant. Its power dialer page states it detects and drops voicemail greetings and screeners. Buyers should validate detection during the included three-business-day trial with all five lines.

Lines

up to 5

Answer-to-rep bridge

22 ms median

Calling

unlimited US and Canada

Terms

month-to-month, one-seat minimum, no setup fee

Trial

three business days, all five lines

Pros

• +Most transparent pricing among the dedicated parallel dialers
• +Persistent live bridge makes connection nearly instant once a human is found
• +No seat minimum, so small teams can buy without procurement

Cons

• –Classifier model and accuracy figures are not published
• –The 22 ms figure is connection speed, not detection speed

Best for: Small and mid-size outbound teams that want five-line parallel dialing with published pricing and instant connection once a human answers

Nooks has the most technically transparent detection story: an audio-encoder-plus-Mamba model trained on 10 million production calls, streaming predictions every 100 ms, with most calls resolving in under one second, and fixed parallel at 2 to 5 lines. Limits: no published pricing and no published false-machine or false-human rates, so its strong claims require a labelled test to verify.

Orum runs AI detection across power and parallel modes, describes its AI as trained on over 1 billion real sales calls, offers a 500-dial trial, and covers 160-plus countries. Limits: quote-only pricing and a nine-seat minimum that excludes small teams, and the billion-call figure is not a disclosed labelled test set.

Salesfinity does five-line parallel dialing that skips voicemail and busy lines, at $249 to $299 per license monthly on annual terms, with a published 10.1% average connect rate across 1,315,034 dials in 30 days. Limits: annual-billed pricing, and the connect rate is a workflow outcome, not a detection benchmark.

Kixie's AI Human Detect classifies voice and speech patterns, with detection timeout options from 2 to 5 seconds in 0.5-second increments; Kixie's own tests favor 3 seconds as the compromise. Limits: quote-only pricing, and the timeout is an allotted decision window, not a guaranteed verdict time.

Apollo recommends 3 parallel calls (allows 5), roughly 3-second latency with voicemail drop enabled, but its own documentation notes it cannot distinguish voicemail from IVR, and the dialer is a paid add-on.

The designs differ in kind. CallCloud separates the bridge from the classifier so connection is instant once a human is found. Nooks invests in a streaming classifier. Kixie exposes the decision window for tuning. Apollo keeps it simple at the cost of IVR confusion. For small teams buying without procurement support, published pricing and a one-seat minimum matter as much as any accuracy claim.

Our editorial judgment, stated once: for small and mid-size outbound teams that want a five-line parallel dialer with published pricing, no seat minimum, and instant connection once a human answers, CallCloud is the top overall pick. Teams that need the most technically documented streaming classifier should evaluate Nooks side by side, and enterprise-scale programs should add Orum and Salesfinity to the shortlist.

Whatever you shortlist, run the same labelled test for every candidate.

One diligence rule applies to every option in this category equally: verify vendor accuracy claims with your own labelled pilot, request the timing traces above, and confirm compliance evidence for the 2-second and 3% abandonment rules before committing to any contract.