Answering machine detection on your own VICIdial server
VM Hunter On-Prem AMD is a Docker container that decides HUMAN or MACHINE from the first two seconds of every answered call, on your hardware. No audio leaves your network, no calls-per-second limit, and a flat yearly price.
The detector sits next to the dialer
The cloud service streams audio to VM Hunter's SpeechLLM engine. On-Prem keeps everything on the server that answers the call: the model, the audio and the verdict.
No audio leaves your network
Detection runs inside a container on your own server. Call audio is analysed there and discarded. The only thing sent to VM Hunter is a call count for billing, every five minutes.
No CPS limit
There is no per-second or per-channel cap. Dial as fast as your carrier and your hardware allow; the license counts calls per year, nothing else.
Decides in two seconds
The model listens to the first two seconds after answer and sets AMDSTATUS and AMDCAUSE for your dialplan, with no round trip to the internet on the call path.
Never blocks dialing
If the container is stopped or the license lapses, the dialplan falls back to the built-in Asterisk AMD() automatically. Your campaigns keep running either way.
Two models, one license
Pick per server. Both ship in the same container and are chosen with one setting; switch any time.
Lite
Runs on the VICIdial server itself
A 0.4 MB acoustic model made for the CPU. It installs on the same machine as VICIdial, so a single-server install needs no extra hardware at all.
- Measured on 16 cores: 5,000 calls in the detection window at once, verdicts in 1.9 s median, 2.9 s p95, no errors; about 2,500 answered calls per second
- A 4-core dialer handles roughly 1,000 at once; 16 GB RAM is plenty for either
- Best for single dialers and most cluster nodes
Full
Most accurate, for high volume
The larger model for the hardest audio: carrier messages, screening prompts and noisy lines. It needs more CPU per call, so size the server for the peak.
- Measured on 16 cores: 1,000 calls in the detection window at once, verdicts in 2.0 s median, 2.6 s p95, no errors; about 500 answered calls per second
- Runs on the dialer or on a dedicated detection box; a GPU lifts it to 5,000 at once
- Best for large campaigns where every verdict counts
| Lite | Full | |
|---|---|---|
| Where it runs | Same server as VICIdial | Dialer or a dedicated server |
| Hardware | 4 cores and up; 8 GB RAM | 16 cores recommended; 16 GB RAM |
| Calls in the window at once, 16 cores | 5,000 (tested limit) | 1,000 |
| Answered calls per second, 16 cores | About 2,500 | About 500 |
| Verdict time at that load | 1.9 s median, 2.9 s p95 | 2.0 s median, 2.6 s p95 |
| Licensed calls per second | No limit | No limit |
| Decision | First 2 seconds of audio | First 2 seconds of audio |
| Audio sent to VM Hunter | None | None |
| GPU | Not needed | Not needed |
Measured with real-time audio streams on a 16-core Ryzen 9 9950X, every call reaching its 2-second decision at the same instant. Older server CPUs do a third to a half of that per core. Figures are for the detector alone; VICIdial and Asterisk need their own headroom on a shared server, so plan on half when they share the box. The install includes scripts/replay.py to measure your own hardware.
Flat yearly pricing by call volume
One license per customer, up to 25 dialer servers, both models, no CPS or channel limits. Renew early and the unused days and calls carry over.
Starts from
for 60 million calls a year. Higher volumes are quoted per customer.
- Both models: Lite and Full
- No calls-per-second limit
- No concurrent-call limit beyond your hardware
- Up to 25 dialer servers per license, more on request
- Verdict from the first two seconds of audio
- HUMAN, MACHINE, SCREENING, SILENCE, TONE and BUSY causes
- Automatic fallback to Asterisk AMD() if the container is ever unavailable
- Works offline for 24 hours to 7 days if the license server cannot be reached
- Model and container updates for the term
- Email support and the install guide
Prices in USD, billed yearly. Tell us your monthly call volume and number of dialers and we reply with a quote the same day.
Up and running in under 30 minutes
The container is a public image. It runs in host-network mode next to Asterisk, listens on a local WebSocket, and the dialplan hands it the call audio through EAGI, the same mechanism VICIdial uses for its own scripts.
- Run the container with your license key:
docker run -d --network host -e LICENSE_KEY=AMD-… ghcr.io/vmhunterhq/acoustic-amd - Add one extension to
extensions.conf: a copy of VICIdial's AMD extension withAMD()replaced by the detector, andAMD()kept as the fallback. - Point your campaign at the new extension. Verdicts land in
AMDSTATUSandAMDCAUSE, so VICIdial's status mapping works unchanged. - Check
docker logs: every verdict is logged with its cause and timing.
The step-by-step guide, the dialplan snippet and a test procedure come with the license. The cloud integration pages for VICIdial and Asterisk describe the same dialplan approach.
Cloud or on-prem?
| VM Hunter cloud | On-Prem AMD | |
|---|---|---|
| Engine | SpeechLLM: understands what the callee says | Acoustic model on the first 2 seconds |
| Where audio goes | Streamed to VM Hunter, logged with transcript | Stays on your server |
| Limits | Calls per month, CPS and channels by plan | Calls per year; no CPS or channel limit |
| Hardware | None | CPU on your dialer (Lite) or a bigger box (Full) |
| Evidence per call | Transcript, recording, cause in the dashboard | Cause and timing in the container log |
| Price | From free, plans from $100 a month | From $2,000 a year |
| Fits | Any dialer, Twilio, voice AI agents | VICIdial and Asterisk sites that must keep audio in-house |
Many customers run both: On-Prem on campaigns with strict data rules, the cloud service where they want transcripts and call-screening detection. The license and the API key are independent.
Questions
What is VM Hunter On-Prem AMD?
What does it send to VM Hunter?
What hardware do I need?
Can it run on the same server as VICIdial?
Is there a calls-per-second limit?
What happens if the license server is unreachable or the container stops?
How is it installed?
How is it priced?
Where do I see usage?
Keep the audio on your servers
Tell us your dialer setup and call volume. You get a license key and the install guide, and the container is live the same day.