Voicemail Detection APIs: Top Options Compared

Written by: Matt Beucler, CEO, Plura AI

Key Takeaways

• A voicemail detection API (AMD) uses audio analysis and machine learning to classify whether an outbound call reached a human or a machine, then returns results via webhook for routing.
• Modern AMD systems combine tone and silence analysis with transcript-based LLM classification, reaching sub-5% false positive rates and sub-second decision times in 2026 benchmarks.³
• Providers differ in granularity: Twilio and Telnyx publish detailed detection types and webhook events, while Vonage, Plivo, and Sinch document only basic human or machine outcomes.⁴
• iOS Live Voicemail and carrier call screening now account for the 30% screening share noted earlier, and Telnyx’s Premium AMD explicitly detects these screeners in real time.
• For high-volume outbound teams, an integrated platform like Plura AI removes the need to stitch together AMD, branded caller ID, TCPA/DNC compliance, and cross-channel memory.²

How Voicemail Detection Works

AMD systems analyze the audio that arrives immediately after a call connects. Traditional approaches examine call progress tones (CPT), silence intervals, and greeting length to distinguish a live human from a machine. Modern implementations layer machine learning classifiers on top of that signal. The most current systems add transcript-based LLM classification to handle ambiguous cases.

Detection result types vary by provider, but common values include:

• human – a live person answered
• machine_start – a machine greeting has begun
• machine_end_beep – the voicemail greeting ended with a beep, ready for a message
• machine_end_silence – the greeting ended in silence
• machine_end_other – the greeting ended in an unclassified way
• IVR – an interactive voice response system answered
• unknown – the system could not classify the pickup

The detection process follows a consistent pattern:

1. Call initiation: Your application places an outbound call with AMD enabled and passes configuration parameters to the provider API.
2. Audio analysis: The provider classifier listens to the first two to eight seconds of connected audio and analyzes speech patterns, silence cadence, and greeting characteristics.
3. Webhook delivery: The provider posts the detection result to your callback URL, and your application executes the appropriate branch. It either connects an agent, plays a voicemail drop, or terminates the call.

A 2026 arXiv preprint (arXiv:2604.09675) validated a production voicemail detection system across 77,000 calls. It achieved a 0.3% false positive rate and a 1.3% false negative rate, with end-to-end inference completing in 46 ms on a commodity dual-core CPU. The research found that temporal speech patterns are the strongest signal for distinguishing voicemail greetings from live human answers.

Top Voicemail Detection APIs Compared

The table below compares the five major CPaaS (Communications Platform as a Service) providers on documented AMD capabilities. Every data point comes from official provider documentation or release notes. Where a provider does not publish a specific value, the table states “not documented.”

Provider	Detection Types	Webhook / Event Names	Notable Features
Twilio	human, machine_end_beep, machine_end_silence, machine_end_other, fax, unknown	AnsweredBy parameter in call status callback	$0.0075 per call, async mode available, DetectMessageEnd mode waits for beep before triggering callback, accuracy claims limited to US and Canada, with UK in development and lower international accuracy
Vonage	human, machine (not documented at granular level)	Callback URL event on detection result	Bundled into voice workflow, no standalone AMD fee documented as of July 2026, billed via per-minute voice charges
Telnyx	human, machine, iOS Call Screening / iOS Live Voicemail (Premium tier)	Webhook on call control event, answering_machine_detection set to “premium” in API call	Premium AMD detects iOS Live Voicemail in real time during active calls, ML-based, recommended for production use
Plivo	human, machine (not documented at granular level)	Not documented at event-name level	Not documented beyond basic human or machine classification
Sinch	Not documented at granular level	Not documented at event-name level	Not documented beyond basic AMD availability

These CPaaS APIs work well for teams that want direct control over their telephony stack. The integration work required to go from a raw AMD webhook to a production-grade outbound system is substantial. You still need to handle branded caller ID, TCPA compliance, DNC scrubbing, stateful conversation memory across channels, and the growing complexity of iOS and carrier-level call screening.

For teams that want those capabilities without building each layer independently, an integrated platform like Plura’s AI predictive dialer delivers voicemail detection as part of a carrier-grade, compliance-supporting whole.

Ready to see how an integrated AMD solution performs at scale? Watch a live walkthrough of Plura’s dialer.

Voicemail Detection API Python Example: Implementation Guide

The following example uses Telnyx’s Programmable Voice API with Premium AMD enabled. It initiates an outbound call with voicemail detection and handles the webhook callback in a Flask application. If you prefer to build on a raw CPaaS API instead of an integrated platform, this pattern shows the core wiring.

import telnyx from flask import Flask, request, jsonify app = Flask(__name__) telnyx.api_key = "YOUR_TELNYX_API_KEY" def make_outbound_call(to_number, from_number, webhook_url): """Initiate an outbound call with Premium AMD enabled.""" try: call = telnyx.Call.create( connection_id="YOUR_CONNECTION_ID", to=to_number, from_=from_number, answering_machine_detection="premium", answering_machine_detection_config={ "after_silence_millis": 1200, "between_words_silence_millis": 100, "maximum_word_length_millis": 3500, "silence_timeout": 5000, "total_analysis_time_millis": 5000, "greeting_duration_millis": 1500, "greeting_silence_duration_millis": 800, }, webhook_url=webhook_url, ) return call except telnyx.error.TelnyxError as e: print(f"Call initiation failed: {e}") raise @app.route("/webhook", methods=["POST"]) def handle_webhook(): """Handle Telnyx call control webhooks including AMD results.""" payload = request.json event_type = payload.get("data", {}).get("event_type", "") call_control_id = payload.get("data", {}).get("payload", {}).get("call_control_id") if event_type == "call.machine.detection.ended": result = payload["data"]["payload"].get("result", "unknown") handle_amd_result(call_control_id, result) elif event_type == "call.answered": # AMD result may arrive separately; hold until detection completes print(f"Call answered: {call_control_id}") return jsonify({"status": "ok"}), 200 def handle_amd_result(call_control_id, result): """Route the call based on AMD classification.""" if result == "human": # Connect to a live agent or start AI conversation print(f"Live human detected on {call_control_id}. Connecting agent.") # telnyx.Call.transfer(call_control_id, to="agent_sip_uri") elif result in ("machine_end_beep", "machine_end_silence"): # Leave a voicemail message print(f"Voicemail detected on {call_control_id}. Leaving message.") # telnyx.Call.speak(call_control_id, payload={"payload": "Hi, this is..."}) elif result == "ios_screen": # iOS Live Voicemail or Call Screening detected print(f"iOS screening detected on {call_control_id}. Hanging up.") # telnyx.Call.hangup(call_control_id) else: # unknown or unhandled result - default to human branch print(f"AMD result '{result}' on {call_control_id}. Defaulting to human branch.") if __name__ == "__main__": app.run(port=5000) 

Common pitfalls to address in production:

• Timeouts: AMD analysis has a maximum window. If the provider returns “unknown” because the greeting was ambiguous, always define an explicit fallback branch. Defaulting to the human branch is usually safer than hanging up on a live prospect.
• Synchronous vs. asynchronous tradeoff: Synchronous AMD blocks the audio pipeline for two to four seconds while analysis runs, which can cause live humans to hang up during the silence. Asynchronous AMD connects immediately and delivers the result via webhook while audio flows. Your application must handle the case where the agent has already started speaking before the voicemail classification arrives.
• Retries: Implement exponential backoff on webhook delivery failures. A missed AMD result that defaults to no action wastes the entire call attempt.
• DetectMessageEnd mode: When leaving a voicemail message, Twilio’s DetectMessageEnd mode waits for the greeting to finish before triggering the callback. This prevents your message from starting mid-greeting. Test both modes against your target carrier mix.

Test both synchronous and asynchronous modes against your target carrier mix. One of the biggest challenges your AMD implementation will face is the rise of iOS Live Voicemail and carrier-level call screening.

Handling iOS Live Voicemail And Call Screening With AMD

The binary human-or-machine classification that defined AMD for a decade now misses too many real-world pickups. A LiveKit Agents user running thousands of answered calls per day reported that roughly 30% of calls were voicemail or iOS 26 Call Screening bots, a share that was climbing as more callers enabled screening.

Three specific challenges now affect every outbound operation:

• iOS Live Voicemail: Apple’s iOS Live Voicemail intercepts calls from unrecognized numbers, transcribes the caller’s message in real time, and lets the recipient decide whether to pick up. The call appears “answered” to your AMD system, but no human is on the line. Standard AMD classifiers trained on traditional voicemail greetings misclassify this as a human pickup, which causes your AI agent to deliver a full pitch to a screening bot.
• Carrier-level call screening: Carriers label high-volume outbound numbers as “Spam Likely” or “Scam Likely,” which routes calls directly to voicemail before they ring. STIR/SHAKEN (Secure Telephone Identity Revisited / Signature-based Handling of Asserted information using toKENs) authentication, which verifies caller identity at the carrier level, reduces the likelihood of low-attestation labels that suppress answer rates.
• Google Pixel Call Screen: Google’s on-device screener answers the call with a synthetic voice, asks the caller to state their purpose, and presents a transcript to the user. Like iOS Live Voicemail, this appears as an answered call to standard AMD.

Telnyx’s Premium AMD, as documented in its May 18, 2026 release notes and developer documentation, detects iOS Call Screening including iOS Live Voicemail in real time during outbound calls, though the evidence does not establish whether other major CPaaS providers offer a comparable feature.

Practical mitigation steps for any AMD implementation work best as a connected strategy. Use branded caller ID so your company name appears on the recipient’s screen, which reduces the likelihood that the call triggers a screening flow in the first place. Pair that with STIR/SHAKEN authentication on every outbound call to maintain high attestation scores with destination carriers.² When a screener is still detected, configure your AMD to hang up cleanly rather than deliver a full pitch to the bot. Finally, test against real iOS and Android devices with screening enabled, not just carrier voicemail systems, because only real-device tests reveal how your audio and timing behave under screening conditions.

Plura addresses branded caller ID and STIR/SHAKEN at the carrier level through its own FCC-licensed infrastructure, issuing branded caller ID directly rather than through a third-party reseller.² This structure differs from CPaaS-wrapper platforms that cannot issue caller ID under their own carrier identity.

See how Plura handles call screening at the carrier level. Request a carrier-level demo.

How To Choose A Voicemail Detection API

The right AMD API depends on your use case, volume, and how much integration work your team can absorb.

For AI voice agents running at low to moderate volume, a transcript-based LLM classifier integrated into your agent framework often delivers the most accurate results. LiveKit’s AMD classifies calls into human, voicemail, IVR, and unavailable, with auto-IVR navigation available in Python. This approach avoids the synchronous blocking delay of traditional CPaaS AMD and handles ambiguous cases through LLM reasoning.

For contact centers and sales dialers at moderate volume, Twilio’s AMD at $0.0075 per call is the most documented option with the widest ecosystem support. Telnyx’s Premium AMD is the stronger choice when iOS Live Voicemail detection is a priority, given its documented real-time iOS screening detection released in May 2026.

For high-volume outbound operations where voicemail detection is one component of a broader compliance, caller ID, and conversation management challenge, standalone CPaaS AMD APIs require significant integration work to reach production quality. You would need to build and maintain branded caller ID, TCPA compliance, DNC scrubbing, stateful conversation memory, and cross-channel context as separate layers.²

Plura’s AI predictive dialer integrates voicemail detection into a carrier-grade platform that also handles branded caller ID issued at the FCC-licensed carrier level, TCPA and DNC compliance support, and stateful conversation memory across voice, SMS, RCS, and webchat.² For operations running thousands of calls per day, the build-versus-buy math often shifts toward an integrated platform. Compare plans and rates to see where the economics land for your volume.

Voicemail Detection API FAQ

What Is the Difference Between Voicemail Detection and Answering Machine Detection?

The terms appear interchangeably in most provider documentation, but there is a technical distinction. Voicemail detection refers specifically to identifying whether a call was answered by a voicemail system. Answering machine detection (AMD) is the broader category that includes voicemail systems, IVR menus, call screening bots like iOS Live Voicemail and Google Pixel Call Screen, fax machines, and other non-human pickups. Modern AMD implementations classify calls into multiple outcome types rather than a simple voicemail-or-not binary, which matters for routing logic in AI voice agent deployments.

How Accurate Are Voicemail Detection APIs?

Published accuracy figures vary significantly by provider, dataset, and measurement methodology. Traditional tone-cadence AMD reaches roughly 85 to 90% accuracy in typical deployments. Machine learning classifiers report false positive rates under 5% in 2026 benchmarks.³ Transcript-based LLM classifiers like LiveKit’s AMD report macro F1 scores around 97% on internal benchmarks.

The most important metric for most operations is the false positive rate, which is the share of live human answers incorrectly classified as voicemail, because each false positive represents a lost conversation. Run your own benchmarks on your specific carrier mix, geographic footprint, and call list before committing to a provider’s published figures.

Can Voicemail Detection Work with AI Voice Agents?

Voicemail detection works well with AI voice agents when you choose the right integration pattern. Synchronous AMD blocks the audio pipeline for two to four seconds while analysis runs, which causes live humans to hang up during the silence. For AI voice agents, asynchronous AMD is usually the preferred approach. The agent connects immediately and the classification result arrives via webhook while audio flows. The agent holds its opening statement in a gate until the AMD result confirms a human, then speaks without delay. Some agent frameworks, including LiveKit Agents, run AMD detection outside the main agent loop so easy cases resolve quickly and only ambiguous ones incur the cost of an LLM call.

How Do I Handle iOS Live Voicemail?

iOS Live Voicemail intercepts calls from unrecognized numbers and presents a synthetic greeting that standard AMD classifiers may misidentify as a human pickup. One direct mitigation uses a provider with explicit iOS screening detection, such as Telnyx’s Premium AMD, which detects iOS Call Screening in real time during active calls. At the carrier level, branded caller ID reduces the likelihood that iOS routes your call to Live Voicemail in the first place, because recognized callers are less likely to trigger the screening flow. STIR/SHAKEN authentication also improves caller ID presentation on iOS devices. When a screener is detected, the preferred behavior is to hang up cleanly rather than delivering a pitch to the bot.

What Is The Best Voicemail Detection API For High-Volume Outbound?

For high-volume outbound operations, the AMD API itself is rarely the binding constraint. The harder problems are branded caller ID, TCPA and DNC compliance, stateful conversation memory, and the operational overhead of maintaining each layer independently. Standalone CPaaS AMD APIs from Twilio, Telnyx, Vonage, Plivo, and Sinch provide solid building blocks, but they require significant integration work to reach production quality at scale.

Plura’s AI predictive dialer integrates voicemail detection into a platform that handles branded caller ID at the FCC-licensed carrier level, supports TCPA and DNC compliance workflows, and maintains cross-channel conversation memory. For teams running thousands of calls per day, that integration reduces both engineering overhead and operational risk compared to assembling the stack from individual CPaaS components.

Conclusion: Build Vs. Buy For High-Volume Outbound

AMD now sits at the core of outbound operations at scale. Forty to sixty percent of outbound AI calls end in voicemail³, and without accurate detection, that volume either wastes agent time or burns per-minute spend on one-sided calls. The CPaaS AMD APIs from Twilio, Telnyx, Vonage, Plivo, and Sinch each offer a documented path to basic voicemail detection, with Telnyx’s Premium AMD standing out for its real-time iOS screening detection.

The build-versus-buy decision turns on how much of the surrounding stack your team is prepared to own. Standalone AMD APIs require you to build and maintain branded caller ID, compliance enforcement, stateful conversation memory, and iOS screening handling independently. For teams with the engineering capacity and the time, that path can work. For high-volume outbound operations where every percentage point of connect rate and every compliance gap carries real cost, an integrated platform delivers those layers as a unit.

Plura’s AI predictive dialer handles voicemail detection as part of a broader platform that includes branded caller ID issued through Plura’s own FCC-licensed carrier, STIR/SHAKEN caller ID verification, TCPA and DNC compliance support, SOC 2, HIPAA, and ISO-related controls, and stateful conversation memory across voice, SMS, RCS, and webchat.¹ Compare plans and rates on Plura’s pricing page, or schedule a live demo to see the full stack in production.

---

Read Next

• Best Voicemail Detection APIs and Dialers for 2026
• What Is Voicemail Detection in Voice AI?
• Voicemail Detection vs Answering Machine Detection
• Voicemail Detection Software Reviews: 2026 Buyer’s Guide
• Voicemail Detection Alternatives for Outbound Leaders