FAS turns call completion into a margin question
False answer supervision can make an unanswered call look billable; tracing signaling, audio and CDRs helps separate carrier costs from AMD mistakes.
The choice is less about where AMD runs than where you can afford its compute, latency and operational complexity.
Outbound dialer teams weighing native Asterisk AMD against an external audio service are choosing where detection work runs, and who has to operate it. The built-in app_amd path keeps the work in Asterisk. A WebSocket-based path sends audio to an external detection engine and returns a verdict through an API. Neither design wins by definition. The right choice depends on call density, measured classification quality and the team’s ability to support the whole path.
Asterisk’s native app_amd returns one of three verdicts: HUMAN, MACHINE or NOTSURE. Its standard approach relies on silence patterns. That makes it a direct option for teams that want AMD close to existing call control and do not want to operate a separate audio-analysis service.
The trade-off is that detection shares the Asterisk environment with telephony work. At high call density, teams should measure CPU use and call handling together rather than assume the module’s footprint is negligible. A change that looks small per call can matter when many channels are active. Test with the codecs, call patterns and concurrency your dialer actually uses.
Classification quality also needs to be measured against real answered calls. Silence-based logic can struggle when a human’s first words are brief, delayed or quiet, and when a recording speaks immediately. A native path may still be the right operational choice, but its verdicts should be audited against call outcomes rather than trusted because they are built in.
With external AMD, audio leaves the call-control process for analysis by another service. A WebSocket API is one way to connect a dialer to that service. This can separate audio analysis from Asterisk’s call-control work and give operators a distinct place to scale or update detection. It also adds a transport path, another service to monitor and another point where delay or failure can affect the call.
“WebSocket audio streaming SIP” can sound like a choice of signaling protocol, but the architecture question is narrower: SIP and dialer call control can remain in place while audio is delivered to an external detector over an API. The integration still needs an explicit plan for audio format, connection handling, verdict timing and what the dialer does if the service is unavailable. Those details vary by implementation; they should be confirmed and tested, not assumed.
External processing does not automatically mean lower CPU use or better accuracy. It shifts work and adds network and service overhead. Measure Asterisk CPU, external-service capacity, stream volume and end-to-end time at peak concurrency. Include failure behavior in the test: a detector that returns a useful verdict quickly in normal conditions is not enough if a broken connection leaves answered calls waiting.
Compare the two approaches using the same call set and a clear definition of error. Count live people classified as machines, machines classified as people, and uncertain outcomes. Also track how long a call takes to reach a usable verdict and how often the dialer cannot act on one. A single accuracy percentage can hide different costs: a false positive can discard a live answer, while a false negative can send an agent to voicemail.
False answer supervision is a separate failure mode. A carrier may report an answer even when no called party answered; changing the AMD classifier does not, by itself, establish what happened in the signaling path. Keep call records and audio evidence in the investigation. The earlier guide to separating FAS from AMD errors covers why those cases need distinct treatment.
Native app_amd suits teams that value a compact Asterisk deployment, can accept its classification behavior after testing, and prefer not to add a separate streaming service. It is also a reasonable baseline for a controlled comparison. Tune and evaluate it against the same operating conditions the production dialer will face.
An external engine suits teams that need to separate analysis from call control, have a clear way to stream audio, and can own the added service and network dependencies. AMDY.IO offers answering-machine detection through native integrations with Asterisk, ViciDial and FreeSWITCH, or through a WebSocket API. It says detection starts in one-eighth of a second and reports 99% accuracy; those are vendor claims, not a substitute for a test on a dialer’s own call mix. Its stated classifications include answering machines, live humans and carrier false answers.
For teams considering that separation, this stack discussion on keeping media handling outside Asterisk provides relevant deployment context. The practical comparison is not “C module versus API” in isolation. It is the total cost of running each path: Asterisk headroom, stream and service reliability, verdict delay, maintenance effort and the errors that matter most to the operation.
Run both options with representative traffic and an agreed scoring method. Record resource use at peak load, end-to-end decision time, failure rates and human-versus-machine errors. Include carrier false answers as a separate category. Then make the choice that fits the team’s capacity to operate it—not the one that sounds simpler on a diagram.
False answer supervision can make an unanswered call look billable; tracing signaling, audio and CDRs helps separate carrier costs from AMD mistakes.
Separate audio analysis from call control: receive the detection result, then use an asynchronous ESL handler to transfer the live channel.
Start audio analysis as soon as the answered channel is available, then route on a returned verdict instead of waiting out a fixed greeting timer.