Compare
AMDSure vs AMDY.io
The closest thing to AMDSure: a model on a server, installed on a VICIdial or Asterisk box with a script or a module, priced by detections. The differences are where the audio goes, how it is secured, and how fast the answer comes.
| AMDSure | AMDY | |
|---|---|---|
| The 3% abandon rule | Shadow mode measures people wrongly sent to voicemail on your own calls, beside your built-in AMD(), with the sample size: your detector's share of the 3% abandon rate, before you switch. People are decided first, 2.55 s on average. | Per-class counts and a sensitivity dial. No side-by-side against your own AMD() on the same calls, and no published error rate on people with a sample size. |
| How it decides | A compiled agent on your Asterisk box measures the call (a real voice detector, tones, a learned acoustic model, the sound of the pickup itself); AMDSure's server applies the rules and signs the verdict. | Your box streams the call's raw 8 kHz audio to AMDY's server over a WebSocket; a model there classifies it into machine types (default greeting, number read-out, Google Voice, iPhone and Samsung screening, no audio, cannot take the call). |
| Where the audio goes | Stays on your server. The agent sends numbers (speech probabilities, tone events, model scores), never audio, unless you switch learning on. | Leaves your server: the raw call audio is streamed to api.amdy.io on port 2700 over a plain (unencrypted) WebSocket, with the API key in a header. A TLS option exists in the native module but is off by default and the documented endpoint is ws://. |
| Time to a verdict | Median 4.6 s to a verdict on live calls from the pickup, 2.3 s on recordings that start mid-greeting; short greetings decide at about 2 s, beeps and known greetings sooner. | Their documentation: typically 5 to 6 seconds after the first sound; final answer once 4.5 s of speech is heard. Our benchmark on the same recordings: median 5.1 s on live clips, 5.0 s on mid-greeting clips. |
| Price | Free: 50,000 verdicts a month, two servers, no card. Starter $69 a month for 500,000 then $0.0002; Growth $269 for 5,000,000 then $0.00012; Scale $899 for 25,000,000 then $0.00008. | Sandbox $0 for 50,000 detections on one server; Starter $79 for 500,000 then $0.00025; Growth $299 for 5,000,000 then $0.00015; Scale $999 for 25,000,000 then $0.00010 (their pricing page, October 2026). |
| What it reports | Person, voicemail, call screener, unsure, each with a reason, a confidence and the time to the verdict. And what answered, by name: Verizon, AT&T and other carriers' default greetings, number read-outs, iPhone, Samsung and Google call screening, a disconnected number (the intercept tones), fax, busy, dead air, a false answer. Stamped on the call as AMDSURE_KIND for the dialer. | HUMAN or an AMD class with a carrier or machine type, plus a duration and confidence string; honeypot and litigator lookups on the phone number as an add-on. |
| Works on | Asterisk 18 to 22 and everything built on it: VICIdial, ViciBox, FreePBX, Issabel, GoAutoDial, custom dialers. FreeSWITCH in development. | VICIdial and ViciBox, Asterisk, FreeSWITCH (mod_audio_fork), and a raw API. |
| Proof before you switch | Shadow mode runs beside your built-in AMD() on every call and stamps its verdict on the call; the dashboard shows people wrongly sent to voicemail and voicemails caught, against AMD(), before you switch. | A dashboard with per-class counts and a 1 to 9 sensitivity. No side-by-side against your own AMD() on the same calls. |
| When it fails | Open, to your own AMD(). A brain outage costs seconds, never a call. | To HUMAN: any error routes the call to an agent. |
| Your audio and training | Your call audio is used for training only with the learning checkbox on, which you can turn off at any time; phone numbers are never seen. | Not stated per call; the service receives the audio of every call. |
Where AMDY is good
Hosted AI detection for VICIdial, Asterisk and FreeSWITCH.
- Easy one-line install on VICIdial.
- Names many carrier greetings by type.
- A long list of analytics reports.
Where AMDSure differs
- AMDY's comparison table says "99% accuracy" and "starts in 1/8 s". No method or sample size is published for the 99%, and the eighth of a second is when listening starts: AMDY's own documentation puts the verdict 5 to 6 seconds after the first sound. AMDSure publishes each error rate with its sample size and the measured time to a verdict.
- AMDY lists "carrier false-answers: classified". AMDSure reports a false answer as its own kind on the call, and a disconnected number (the intercept tones) too, which AMDY does not name.
- AMDY detects honeypots by looking the phone number up on a list. AMDSure never sees a phone number, by design, so that stays a dialer feature; it is the reason our audio and data story is simple.
- AMDSure sends numbers, not audio, and everything is over TLS; AMDY streams the call itself, unencrypted, with a permanent API key.
- On the same 18 labelled live calls AMDSure was right on 15 and AMDY on 14, at 4.6 s against 5.1 s median; on 173 older mid-greeting recordings 109 against 113, at 2.3 s against 5.0 s. AMDY is stronger on voicemail greetings with the opening missing; AMDSure on people, dead air and screeners.
- AMDSure falls back to your own AMD(), not to HUMAN, so an outage costs seconds rather than agent time on machines.
- AMDSure runs in shadow beside your AMD() and shows both on your calls; AMDY's installer replaces the AMD step.
- AMDSure never sees a phone number, so honeypot and litigator lists are a dialer feature, not ours.
Sources
Facts about AMDY come from its own documentation and pricing pages on the dates noted; tell us at [email protected] if something has changed and we will fix it. Other comparisons: Stock AMD(), FreeSWITCH AMD, Twilio AMD, Telnyx AMD, Plivo, SignalWire AMD, Vonage, VM Hunter.
See it on your own traffic.
A 20 minute call. We show the console with live numbers, then set up shadow mode on your dialer so you can judge for yourself.
Start free