Silent treatment: how call screening broke legacy dialer AMD

The breakdown of silence-based detection

Predictive dialers have relied on basic audio detection algorithms for over twenty years. Asterisk applications like app_amd listen for energy thresholds and pause lengths. They count milliseconds of silence, estimate speech duration, and make a binary guess: human or machine. When open-source platforms like ViciDial or FreeSWITCH make outbound calls, this silence-counting logic determines whether to hand off the SIP session to an agent or drop the channel.

This approach worked when outbound targets were mostly landlines and traditional carrier voicemails. It fails on mobile networks. Modern smartphone features, particularly iOS Live Voicemail and Android call screening, have rendered silence detection unreliable. When an iPhone intercepts an outbound call, it answers immediately, plays a variable text-to-speech prompt, and displays a live transcript to the recipient. The silence detection algorithm sees an immediate answer followed by structured audio. It misidentifies the prospect as a voicemail machine and terminates the call.

In practice, standard dialer AMD misclassifies an estimated 10% to 20% of live answers as machines. When an engine makes a false positive call, the dialer hangs up or routes the call away. The prospect picks up, hears silence or a click, and adds the calling number to a spam registry. For operations running high-volume outbound campaigns, these false positives destroy valid leads and degrade caller ID reputation rapidly.

How silence algorithms fail under smartphone screening

To understand why legacy tools fail, examine how standard app_amd processes incoming media streams. The dialer opens an analysis window only after receiving a SIP 200 OK signal. The algorithm tracks three variables: initial silence before speech, length of the first greeting, and silence trailing the greeting.

• Initial silence threshold: If a human answers with a delayed "hello," standard AMD often interprets the initial pause as an answering machine beep delay.
• Greeting length limits: Human answers are short and ragged, usually under 1.5 seconds. Smartphone call screening prompts last several seconds. Legacy AMD flags any greeting longer than two seconds as a recorded machine greeting.
• Silence cadence: Carrier voicemails play a continuous prompt followed by a sharp tone. Smartphone screening prompts introduce variable pauses while transcribing speech, confusing basic audio energy detectors.

We detailed these mechanics previously when comparing outbound answering machine detection in ViciDial and Asterisk. When silence counters miscalculate, the result is either a false positive—hanging up on a live prospect—or a false negative, which routes recorded audio to an agent. Neither outcome is acceptable in an efficient call center.

Real-time audio stream analysis vs legacy AMD

Machine learning models approach audio detection differently. Rather than counting silence gaps on a timer, stream-based detection processes raw RTP audio packets directly. Systems like AMDY.IO analyze acoustic fingerprints rather than relying purely on speech cadence or silence duration.

Stream analysis begins evaluating audio within 1/8 of a second of the answer signal. By inspecting the frequency distribution, background noise, and acoustic signatures of the first few milliseconds of media, the engine distinguishes between a human vocal tract, a recorded greeting, and an iOS Live Voicemail prompt.

This structural difference alters detection performance across several core operating metrics:

1. Detection latency and dead air

Legacy silence algorithms must wait for a speaker to finish talking—or for a silence threshold to expire—before returning a verdict. This introduces two to four seconds of delay. During this window, the called party hears nothing. This dead air triggers immediate hangups and creates compliance risks under telemarketing sales rules. We covered this dynamic in our report on how AMD latency triggers TCPA abandonment rules. Stream-based analysis makes classification decisions almost instantly, allowing the dialer to bridge the agent before the prospect finishes saying "hello."

2. Accuracy on call screening

Because audio fingerprinting evaluates spectral properties rather than length alone, it correctly identifies smartphone screening prompts. A live human who pauses before speaking is recognized as human. An automated transcription prompt is recognized as an automated system, even if its pause duration mimics human speech.

3. Carrier False Answer Supervision (FAS)

Carrier false answers occur when intermediate carriers return a SIP 200 OK signal without establishing a connection to the far end. The dialer begins billing and timing the call while audio stream analysis detects total silence, dead air, or carrier intercept tones. Legacy AMD treats FAS as a normal call with long initial silence, burning dialer capacity. Audio stream systems flag FAS immediately, preventing unnecessary carrier billing and inaccurate call logging.

Choosing the right engine for your dialer stack

Selecting between legacy silence detection and machine learning stream analysis depends on campaign type, dialer architecture, and call volume.

When legacy silence AMD is sufficient

Stock app_amd or basic dialer settings remain functional for small-scale operations with specific parameters:

• Low call volume: Teams running single-line manual dialing or low-ratio power dialers where agent idle time is not critical.
• B2B landline targeting: Outbound campaigns calling direct office landlines where iOS Live Voicemail and mobile screening features do not exist.
• Zero software budget: Operations that cannot add external services and rely exclusively on built-in Asterisk configurations.

When audio stream analysis is necessary

Stream analysis becomes necessary when dialing consumer mobile numbers or operating predictive dialers at scale:

• High-volume consumer outbound: Campaigns targeting mobile phones where modern call screening causes 10% to 20% lead loss under legacy AMD.
• Predictive dialer efficiency: Operations running high dial ratios on ViciDial, Asterisk, GoAutoDial, or FreeSWITCH that cannot tolerate two-second post-answer delays.
• Compliance management: Call centers needing to eliminate silent calls and stay strictly under TCPA and Ofcom abandonment thresholds.

AMDY.IO provides audio stream analysis with 99% accuracy, starting detection in 1/8 of a second. It integrates natively with ViciDial, Asterisk, FreeSWITCH, and GoAutoDial, or connects via a WebSocket API. The service includes zero setup fees and offers 50,000 free call detections upon signing up. For teams struggling with dropped leads and ghost calls, moving beyond silence duration algorithms is the clearest route to higher connect rates.