Guides
How do parallel dialers detect voicemail versus a live person?
Ring signals, greeting timing, beeps, AI audio models and a human ear. How each one works, why speed costs you real people, and how to test any dialer's detection on your own calls.
Parallel dialers tell voicemail from a live person by analyzing the first moments of audio after a line picks up: how long the greeting runs, where the pauses fall, whether a beep follows, and, in products such as Orum, Nooks and Salesfinity, an AI model trained on recorded calls. Most connect the rep only after the software decides it heard a person. CallCloud works the other way round: it joins the rep to each answered line right away (22 ms median) and lets its detector press Voicemail only where that is safe, so a person never waits on a verdict. Every method trades speed against mistakes, and the costly mistake is hanging up on a real person.
We build CallCloud, one of the dialers compared below, so treat this as a vendor's view. Where another product fits better, we say so.
The five ways a dialer can tell
Most products combine several. None is proof on its own.
1. Ring and pickup signals
Before anyone speaks, the phone network tells the dialer that a number is ringing and then that the call was answered, and call progress analysis can catch disconnected-number announcements and network tones. None of that says who answered: a voicemail system picks up the same way a person does. US rules also require covered telemarketing calls to ring for at least 15 seconds or four rings before the dialer gives up.
2. Greeting length and silence cadence
This is classic answering machine detection (AMD). The software splits the audio into speech and silence and applies timing rules. A person usually says a short "Hello?" and waits; a recording talks for several seconds without a break. Asterisk's open source AMD module lists the rules: initial silence, maximum greeting length, silence after the greeting and a word count (bursts of sound, not recognized words). Twilio documents the two weak spots: a short recorded greeting gets judged a person, and a long human greeting, like a receptionist's, gets judged a machine.
3. Beep detection
The tone at the end of a greeting means the mailbox is recording. It usually arrives too late to decide whether to connect the rep, so dialers use it as a second step that times the voicemail drop. Five9 waits for a beep or 2 seconds of silence, and warns that a pause inside a greeting can start the message too early.
4. AI audio classifiers
Newer dialers feed the audio to a model trained on labeled calls. Nooks has published the most detail: an audio encoder with Mamba layers that makes a fresh prediction every 100 ms. Another design, documented by LiveKit, transcribes the greeting and classifies the words. A model catches patterns that timing rules miss, but it still needs a confidence threshold, so it still makes both kinds of error.
5. A human clicking Connect
The oldest method is a person listening. On a single-line power dialer like PhoneBurner, the rep is on the line from the first word, and PhoneBurner describes its design as having "no voice detection, no transfer, and no pause." ConnectAndSell's Precision Connect uses human agents to get past executive assistants and gatekeepers. A human is hard to fool, but someone has to listen to every answer.
The trade-off: speed against hanging up on people
Every detector makes two kinds of mistake, and they do not cost the same. A voicemail passed to the rep costs a few seconds and a click. A person judged to be a machine is hung up on, or hears silence while the software decides, and that conversation is usually gone.
Listening longer cuts the first mistake and makes the person wait. Twilio's default synchronous AMD averages about 4 seconds, and JustCall says its predictive AMD takes about 4 to 5 seconds to analyze and connect an agent. Tunable products show the same dial: Kixie's timeout runs from 2 to 5 seconds, Readymode offers a 3-second Quick and a 7-second Rigorous profile, and Orum's Fastest mode connects more aggressively than Default and can send reps more voicemail.
That wait is what the US abandoned call rule is about. Under the FCC rule at 47 CFR 64.1200(a)(7), a telemarketing call counts as abandoned if the person is not connected to a live sales representative within 2 seconds after their completed greeting. Abandoned calls may not exceed 3% of calls answered by a live person, measured per campaign over 30 days. The clock starts when the greeting ends, not at pickup, so a vendor's "seconds after answer" figure is a different measure. The FTC's Telemarketing Sales Rule has a related safe harbor; ask counsel whether either covers your calls.
Parallel dialing adds a failure no classifier fixes: two people answer at once and there is one rep. Salesfinity's documentation says the second person can be hung up on while the rep is talking. Ask every vendor what happens to that person.
Which dialer has the best answering machine detection?
Nobody can honestly name one. No vendor, ourselves included, publishes accuracy measured on a shared test set of calls, so the figures cannot be ranked against each other. Convoso claims about 98.9%. JustCall's help center says its AMD is typically 75% or higher. Twilio says one mode with US default settings gets results close to 100%, and its 2017 launch post claimed 94%. A 2012 VICIdial presentation put the Asterisk detector of that time at 70 to 80% at best. Each comes from the vendor's own calls, definitions and clock. Orum, Kixie, Salesfinity, Trellus and CloudTalk publish no percentage, Nooks publishes no single percentage, and Koncert promises "reasonable accuracy" without a number.
One accuracy number also hides the split that matters. A detector that passes many voicemails to reps but never drops a person, and one that blocks nearly every voicemail but drops 1 person in 30, can post similar overall scores. The best detection is the one that makes the fewest of the mistakes you cannot afford, on your own list. You can measure that in a week.
How to test it yourself
- Build a labeled call set. People with short and long hellos (teammates work), office lines with receptionists and phone menus, screened mobiles, short and long voicemail greetings, and full mailboxes.
- Record and label every call yourself. The dialer's own disposition is the thing you are testing.
- Count two numbers separately. People hung up on or left in silence, out of all people who answered; voicemails that reached a rep, out of all voicemails. Ask each vendor for both.
- Time the silence. From each recording, measure the gap between the end of the person's hello and the rep's first word. That is the gap the FCC rule measures.
- Answer two calls at once. Have two teammates pick up within a second of each other and see what happens to the second one.
- Check that drops landed. Listen to the test mailboxes. A "voicemail left" disposition does not prove the whole message arrived.
- Compare conversations, not verdicts. Run a week on your real list and count live conversations per rep hour.
What each vendor publishes
Features, prices and figures as of November 2026, from each vendor's own pages. Read the accuracy column as what each company says about itself, not as a ranking.
| Dialer | How it decides | Does the rep hear the answer live? | Published accuracy | Voicemail drop |
|---|---|---|---|---|
| CallCloud Parallel Dialer | Voice-activity timing, known greetings and a neural audio model trained on its own labeled calls; no transcription service. Auto-presses only on mobile and home lines | Yes. Joined to each answered line at once (22 ms median) while detection runs | No guarantee; one morning of live calls reported below | Yes. Per-rep recordings, unlimited and free, never the same message twice to one person |
| Orum | AI listens for a live voice; Default or Fastest mode, parallel only | After the AI decides; Fastest connects more aggressively and can pass more voicemail | None published | Yes. Automatic recorded drop, or skip, when a machine is detected |
| Nooks | Streaming audio model, a prediction every 100 ms; separate beep detection | Parallel: after the model decides. Power: connected by default | No single percentage | Yes. In parallel, automatic recorded drop when the beep is detected |
| Salesfinity | Lightning AMD, classifying progressively in real time | After detection; claims about 400 ms median (another Salesfinity guide says 600 ms) | None published | Yes. Automatic after the beep or silence; Gold, Enterprise and Enterprise+ seats |
| Trellus | Connects the rep when a person answers; method not described | After detection; claims a 300 ms pickup delay | None published | Yes. On parallel calls it waits for the tone, then plays the recording; included in Power ($99.99 a seat) and Power + Parallel ($149.99) |
| Kixie | AI Human Detect on the PowerDialer, timeout set from 2 to 5 seconds | After detection, within the timeout you set | None published | Yes. Rep triggers it, Kixie waits for the beep. AI detection is a premium add-on |
| CloudTalk | AMD on each line in Parallel mode; unsure results go to the rep | After a short check; CloudTalk says the rep may miss words spoken during it | None published | Yes. Automatic in Parallel mode, manual in sequential |
| Koncert | AI recognizes voicemail, phone trees and people, in AI Parallel and one-line AI Flow | After the AI decides; in Flow the rep can take control with a button | "Reasonable accuracy," no percentage | Yes. Automatic recorded drop in both |
| Convoso | Words and silence in the opening audio plus a model trained on labeled calls; admin settings | After detection: about 1.4 seconds to decide, then about 200 ms to connect | About 98.9% (vendor claim) | Yes. Automatic, Basic, Advanced or Smart |
| JustCall | AMD in the Predictive Dialer | After detection: about 4 to 5 seconds | Typically 75% or higher | Yes. Rep-triggered recorded drop on all plans |
| PhoneBurner | No detection; one line, and the rep listens | Yes, from the first word | Not applicable | Yes. One-click recorded drop on all plans |
Where another product is the better fit
- You want to tune the detector yourself. Twilio documents four timing settings in its API, and Convoso, Readymode, Five9 and Kixie give admins greeting, silence or timeout controls. Worth it if someone will own the settings.
- You need people to get past gatekeepers. ConnectAndSell's Precision Connect puts human agents in front of executive assistants, a service no self-serve dialer, ours included, replaces.
- You run contact center campaigns at volume. Five9 can play a recorded prompt or a text-to-speech script to machines with no agent involved, and Convoso offers Basic, Advanced and Smart automatic drops.
- You dial one number at a time and want no detection at all. A single-line power dialer such as PhoneBurner keeps the rep on every answer from the first word.
How CallCloud handles it
In the CallCloud Parallel Dialer, the rep is joined to each answered line right away, a measured median of 22 ms (see the zero-lag measurement), and the line shows "Classifying" on screen. The rep hears the greeting live and can press Connect, Voicemail or Skip at any moment, so detection never puts a person on hold.
Alongside the rep, a detector uses voice-activity timing, greetings it has heard before, and a neural audio model trained on our own labeled calls. No transcription service is involved. When it judges that a mobile or home line reached voicemail, it presses Voicemail for the rep, and the campaign's Voicemail setting decides what happens next.
Why office lines stay with the rep
The detector does not press Voicemail on office numbers, because receptionists, phone menus and answering services sound like recordings. In live calls on October 12, 2026, ours wrongly pressed Voicemail on 7 of 19 people who answered office numbers, against 0 of 141 people on mobiles. A wrong press drops a real person; a missed voicemail costs the rep, who is already listening, one click. So on office lines, the call stays with the rep.
What happens to a voicemail
Each campaign has a Voicemail setting with three choices. Skip hangs up. Leave a message plays your voicemail message after the beep while you take the next call. Give to me leaves the call with the rep. Each rep records their own messages, with the campaign's message as a fallback. Drops are unlimited and free, and CallCloud never leaves the same message twice for the same person.
One morning of live calls
On the morning of October 26, 2026, across 1,802 answered calls, 1,109 reached voicemail. The detector pressed Voicemail on 839 of those 1,109, with a median time of 0.48 seconds after the answer. The other 270 voicemails were left for reps to press. Over the same morning, it wrongly pressed Voicemail on 1 of 453 people who answered. That is one morning of live calls, not an accuracy guarantee.
Our limits: the detector misses some voicemails (270 of 1,109 that morning), which reps then press themselves, and office lines always need the rep's judgment. A design that decides before connecting the rep can keep more voicemails away from reps, at the price of the wait described above.
The Parallel Dialer is $125 per seat per month for up to five lines at once; the Power Dialer is $60 for one line. The free trial dials all five: start a trial. For line counts, see power dialer vs parallel dialer.
FAQ
How do parallel dialers detect voicemail versus a live person?
They analyze the first moments of audio after a line picks up: greeting length, pauses, beeps and, in newer products such as Orum, Nooks and Salesfinity, an AI model trained on recorded calls. Most connect the rep only after the software decides a person answered. CallCloud joins the rep to each answered line right away (22 ms median) and lets its detector press Voicemail only on mobile and home lines.
Which dialer has the best answering machine detection?
No one can say from public data, because no vendor publishes accuracy measured on a shared test set. Convoso claims about 98.9% and JustCall says typically 75% or higher, but each figure uses its own calls and definitions. Test dialers yourself on labeled calls and count people wrongly dropped and voicemails passed to reps as two separate numbers.
Can answering machine detection hang up on real people?
Yes. A short human hello can look like a recording and a receptionist's long greeting can look like a machine, so detectors sometimes drop a real person or leave them in silence. In parallel dialing, a second person who answers while the rep is busy can also be hung up on. That is why it matters what a dialer does when it is unsure.
What does the FCC rule say about AMD delay?
Under 47 CFR 64.1200(a)(7), a telemarketing call is abandoned if the person is not connected to a live sales representative within 2 seconds after their completed greeting. Abandoned calls may not exceed 3% of calls answered by a live person, per campaign over a 30-day period. Ask counsel whether the rule covers your calls.
How fast is voicemail detection?
Published figures vary by design and by where the clock starts. Twilio's default synchronous AMD averages about 4 seconds, JustCall says 4 to 5 seconds, Convoso says about 1.4 seconds to decide plus about 200 ms to connect, and Salesfinity claims about 400 ms median. On one morning of live calls on October 26, 2026, CallCloud's detector pressed Voicemail with a median time of 0.48 seconds while the rep was already on the line.
What is a voicemail drop?
A voicemail drop plays a message recorded in advance into the prospect's mailbox, usually after the beep, while the rep moves to the next call. Orum, Nooks, Salesfinity, Trellus, Kixie, PhoneBurner and CallCloud all offer one in some form. Check that the full message lands, since a voicemail-left disposition does not prove delivery.
Why does CallCloud leave office lines to the rep?
Receptionists, phone menus and answering services fool voicemail detectors. In live calls on October 12, 2026, CallCloud's detector wrongly pressed Voicemail on 7 of 19 people who answered office numbers, against 0 of 141 on mobiles. So on office lines the decision stays with the rep, who already hears the call live.
Sources
- support.orum.com/en-US/orum/article/wRiuD9Xx-ai-detection-modes
- support.orum.com/en-US/orum/article/ART-1076
- nooks.ai/blog-posts/low-latency-human-detection-at-scale
- support.nooks.ai/articles/6949493085-automatically-skipping-voicemail-connects
- docs.salesfinity.ai/dialer/parallel-dialing
- docs.salesfinity.ai/dialer/voicemail-and-dispositions
- salesfinity.ai/product/ai-parallel-dialer
- salesfinity.ai/blog/parallel-dialer-without-delay
- trellus.ai/faq
- trellus.ai
- trellus.ai/pricing
- support.kixie.com/hc/en-us/articles/22656918351259-PowerDialer-with-AI-Human-Detect
- support.kixie.com/hc/en-us/articles/4576581944731
- kixie.com/pricing
- help.cloudtalk.io/en/articles/17025675-parallel-mode-2-lines
- koncert.com/hubfs/License/Description%20of%20Koncert%20Products.pdf
- convoso.com/blog/voicemail-detection-boosts-contact-rates
- help.convoso.com/hc/en-us/articles/43971469658139-Call-Routing-Analysis-Campaign-Settings
- help.justcall.io/en/articles/8559982-understanding-answering-machine-detection-amd-in-predictive-dialer
- help.justcall.io/en/articles/1478436-how-to-set-up-and-use-voicemail-drops-in-justcall
- twilio.com/docs/voice/answering-machine-detection
- twilio.com/docs/voice/answering-machine-detection-faq-best-practices
- twilio.com/en-us/blog/products/launches/introducing-new-answering-machine-detection-html
- twilio.com/en-us/voice/pricing/us
- phoneburner.com/blog/power-dialers-vs-predictive-dialers
- phoneburner.com/faq
- phoneburner.com/pricing
- community.five9.com/s/article/Administrator-Actions-on-Answering-Machine-Detection
- help.readymode.com/support/solutions/articles/11000135400-advanced-machine-detection
- connectandsell.com
- docs.asterisk.org/Latest_API/API_Documentation/Dialplan_Applications/AMD
- docs.livekit.io/telephony/features/answering-machine-detection
- vicidial.org/presentations/MattFlorell_Astricon_2012_CPA-AMD.pdf
- ecfr.gov/current/title-47/section-64.1200
- ftc.gov/business-guidance/resources/complying-telemarketing-sales-rule
