Half-Duplex vs Full-Duplex: The Fundamentals
In telecommunications, duplex refers to how two parties communicate over a channel.
Half-duplex (walkie-talkie): One party transmits at a time. You press to talk, release to listen.
Full-duplex (telephone): Both parties can transmit simultaneously. Normal phone calls work this way.
All voice AI — until very recently — has been half-duplex. And that's not an accident. It's genuinely hard to build a full-duplex AI system.
Why Half-Duplex AI Exists
Traditional voice AI pipelines look like this:
User speaks
→ ASR (speech-to-text)
→ LLM (text reasoning)
→ TTS (text-to-speech)
→ AI speaks
→ User speaks againThis is a sequential pipeline. Each stage must finish before the next starts. And the whole thing resets between turns.
Building truly simultaneous processing requires throwing away this pipeline architecture entirely.
What Full-Duplex Requires
For a voice AI to operate in full-duplex mode, it needs to solve three hard problems simultaneously:
1. Continuous Audio Processing
The model must process incoming audio even while generating output. This means no "recording window" — just a continuous stream.
2. Acoustic Echo Cancellation
When the AI speaks through a speaker, that sound enters the microphone. The model must cancel its own voice from the input signal, or it will hear itself and get confused. This is an engineering challenge as much as an AI one.
3. Turn-Taking Intelligence
The hardest problem: knowing when the human is done speaking. In half-duplex systems, this is solved with silence thresholds. In full-duplex, the model needs real-time understanding of conversational state:
- Is this silence a thinking pause?
- Is the human trailing off to invite a response?
- Is the human trying to interrupt?
- Is that background noise or a new utterance?
How Seeduplex Solves These Problems
Seeduplex's approach is a joint acoustic-semantic model. Rather than separating acoustic processing from language understanding, both happen in the same model, at the same time.
This means the model can use context to inform perception:
- If the conversation suggests the human is mid-thought, it won't misinterpret a pause as an ending
- If it recognizes background speech patterns (navigation, TV), it down-weights those signals
The result is turn-taking that uses both the audio signal and the semantic state of the conversation — not just silence thresholds.
Why This Matters for Real Applications
Customer service: Agents can be interrupted, clarify mid-sentence, and respond to "wait, no, I meant..." naturally.
Voice companions: Conversation feels like talking to a person, not querying a system.
Real-time translation: Both speakers can overlap naturally as translators do.
Accessibility: Users with speech patterns that don't fit standard silence-threshold detection can communicate more effectively.
The State of Full-Duplex AI in 2026
Seeduplex is the first production-scale full-duplex voice model. A few research systems have demonstrated full-duplex capabilities in labs, but deploying at scale (hundreds of millions of users) introduces engineering challenges that research prototypes don't face.
Other major AI labs are working on similar capabilities. The half-duplex era of voice AI is likely ending — Seeduplex is just the first production model to cross the line.