Why AGI Cannot Provide Real-Time Duplex Voice (Technical Limits)

Many Asterisk-based AI calling systems start with AGI (Asterisk Gateway Interface). It is simple, familiar, and easy to attach to a dialplan. But the moment you try to build a natural, interruptible, real-time voice assistant, AGI becomes the bottleneck.

This is not a performance tuning issue. It is not a threading problem. It is not a scaling problem. It is a design limitation.

Real-time duplex voice requires continuous bidirectional RTP media control. AGI was never designed for that.

What Real-Time Duplex Actually Requires

A true duplex AI voice system must:

• Stream caller audio continuously to AI (not in chunks after playback).
• Generate AI speech while still receiving caller audio.
• Detect caller interruptions instantly (barge-in detection).
• Stop playback mid-frame without waiting for file completion.
• Maintain stable RTP timing (20ms cadence, monotonic timestamps).

In protocol terms, duplex means:

• Two active RTP directions.
• No blocking media primitives.
• Application-level control over bridge and media state.

AGI does not provide this control surface.

How AGI Actually Works

AGI is synchronous and command-based.

The call flow looks like this:

Caller → Dialplan → AGI script
AGI sends command → Asterisk executes → AGI waits
  

Each AGI command blocks until completion:

• STREAM FILE blocks until playback finishes.
• RECORD FILE blocks until recording ends.
• GET DATA blocks waiting for input.

While one operation is running, AGI cannot simultaneously process another.

The Blocking Playback Problem

The biggest limitation: STREAM FILE is blocking.

When Asterisk plays audio through AGI:

• The channel enters playback state.
• AGI waits for playback to finish.
• No concurrent audio capture occurs.

That means:

• You cannot detect interruptions during playback.
• You cannot cut speech mid-stream reliably.
• You cannot stream AI-generated audio in small RTP frames.

The system becomes turn-based:

AI speaks → Caller waits
Caller speaks → AI waits
  

This is half-duplex, not real duplex.

No Direct RTP Access

Duplex voice requires tight control over RTP:

• Precise 20ms frame timing
• Stable sequence numbers
• Monotonic timestamp increments
• Payload type consistency

AGI does not expose RTP streams.

It only allows high-level media commands. You cannot inject raw RTP frames. You cannot read live RTP in parallel. You cannot control SSRC behavior.

Without RTP-level access, real duplex media engineering is impossible.

Why Barge-In Fails Under AGI

True conversational AI must support barge-in:

• User interrupts AI.
• AI stops speaking instantly.
• System switches to listening state immediately.

Under AGI:

• Playback blocks execution.
• Input detection happens only after playback completes.
• Speech overlap cannot be resolved cleanly.

The result is unnatural conversation flow.

Latency Amplification

AGI introduces additional latency layers:

• Dialplan → AGI process handoff
• STDIN/STDOUT command parsing
• Command execution round-trips

In AI voice systems, even 200–300ms delays degrade naturalness. AGI makes low-latency conversational timing extremely difficult.

Why ARI + ExternalMedia Solves This

Real duplex requires asynchronous media control.

ARI provides:

• Event-driven channel state control
• Bridge manipulation in real time
• ExternalMedia channels for raw RTP streaming

With ExternalMedia:

• Asterisk sends live RTP to your application.
• Your application sends RTP back concurrently.
• No blocking playback primitives are involved.

That enables:

• True bidirectional streaming
• Instant barge-in detection
• Frame-level playback control
• Stable RTP timing discipline

The Architectural Difference

AGI Model:

Caller ↔ Asterisk ↔ AGI (commands)
  

Media is controlled by blocking instructions.

ARI + ExternalMedia Model:

Caller ↔ Bridge ↔ ExternalMedia ↔ Application
  

Media flows continuously in both directions. Control is event-driven and asynchronous.

Production Reality

Many systems attempt to “extend” AGI for AI. They eventually hit these symptoms:

• One-way audio
• Playback delay
• Missed interruptions
• Conversation overlap artifacts
• High perceived latency

These are not bugs. They are architectural mismatches.

Final Conclusion

AGI was designed for scripted IVR logic, not real-time AI conversation.

Real-time duplex voice requires:

• Non-blocking media handling
• Bidirectional RTP control
• Bridge-level orchestration
• Event-driven state management

Those capabilities exist in ARI + ExternalMedia — not in AGI.

If your goal is a natural, interruptible, human-like AI conversation, AGI is structurally incapable of delivering it.