We released an on-prem answering machine detector that runs on the same box as Asterisk/VICIdial

We run a hosted AMD service (VM Hunter). The most frequent request from VICIdial and Asterisk operators was for an option that does not stream audio to a third party. So we built one.

It is an acoustic model (CNN on log-mel spectrograms plus a call-progress tone detector) served locally over a WebSocket. The Asterisk side is a static C EAGI binary with no dependencies that sets AMDSTATUS/AMDCAUSE, replacing AMD() in the VICIdial AMD extension. Verdict at 2.0 s. Lite model does ~5,000 concurrent streams on a 16-core CPU.

Honest limits: it is weak on calls where the first word is a lone "Hello", because a person and a greeting sound the same for two seconds. We published a 4,000-call comparison against our speech-based hosted detector, including the failure cases: https://vmhunter.com/blog/acoustic-amd-vs-speech-amd-4000-call-test

The model, training code, and client are MIT-licensed on GitHub (github.com/vmhunterhq/vmhunter-acoustic-amd). We're announcing today the supported, licensed build with the portal and updates. Happy to answer anything about the architecture or the test.