A caller pauses to think for half a second and the agent jumps in anyway. That single design flaw is the difference between an agent that feels broken and one that feels like a real conversation, and it comes down to which of three architectures is running underneath the call.
Full-duplex voice AI is an architecture that processes incoming audio and generates outgoing audio at the same time, deciding many times per second whether to speak, keep listening, pause, or hand a question off to another model, instead of waiting for a gap in speech to decide a turn has ended. It is the newest of three approaches, after cascaded and turn-based systems, and it is the architecture behind OpenAI's GPT-Live, which made the term mainstream on July 8, 2026.
Most production voice agents today, including the majority of what runs on Vapi, Retell, and Bland, are not full-duplex. They are cascaded. That gap between what just got announced and what is actually deployed is worth understanding before it shapes a client conversation.
What Is the Difference Between Cascaded and Full-Duplex Voice AI?
A cascaded system chains three separate models: speech-to-text transcribes the caller, a language model generates a text response, and text-to-speech converts that response back into audio. Each step waits for the one before it to finish. It is modular (swap a transcription vendor without touching the rest of the stack), and it is what most phone-based agents run today, largely because it lets a team plug in a transcription model with years of production tuning on narrowband phone audio, not because full-duplex models are structurally unable to process a phone line.
A full-duplex system skips the handoffs. One system listens and generates continuously, which is what allows it to backchannel naturally, hold a real pause without jumping in, and recover gracefully when the caller interrupts mid-response. GPT-Live is the clearest public example: it drops in "mhmm" or "yeah" while the caller is still talking, waits instead of interrupting when someone pauses to think, and OpenAI says it now filters out background noise, like traffic or a nearby conversation, better than the turn-based version it replaced. Those are the exact details non-technical buyers point to when they say an agent "feels human" after watching a demo.
The tradeoff is that a full-duplex system is closer to a black box: harder to inspect, harder to insert a compliance filter into a specific step, and not yet proven at scale over compressed telephony audio.
Where Does Turn-Based Voice AI Fit Between the Two?
Turn-based models, like the version of ChatGPT Advanced Voice Mode that ran on GPT-4o, sit in the middle. They collapse the pipeline into one model handling audio in and out, which removes the handoff delay of a cascaded system and produces a noticeably smoother response.
The limitation is in the name. A turn-based model still operates on discrete turns: it waits for silence, interprets that silence as the end of the caller's turn, and then responds. Because turn detection depends on silence rather than genuine understanding of conversational flow, a caller pausing to think, coughing, or getting drowned out by background noise gets misread as done talking, and the model jumps in at the wrong moment.
| Cascaded | Turn-based | Full-duplex | |
|---|---|---|---|
| How it processes audio | Sequential: STT, then LLM, then TTS | One model, audio in and out | Continuous, simultaneous in and out |
| Turn detection | Explicit step between STT and LLM | Silence-based, single model | Continuous decision, many times a second |
| Track record on narrowband phone audio | Extensive; years of telephony-specific tuning | Limited; mostly app and web deployments | None yet at production scale over phone lines |
| Can you inspect and filter each step | Yes | Limited | Largely a black box |
| Example | Original ChatGPT Voice, most Vapi/Retell/Bland deployments | ChatGPT Advanced Voice Mode (GPT-4o) | OpenAI GPT-Live |
Why Do Most Production Phone Agents Still Run Cascaded Architecture?
Phone audio is a genuinely harder input than the wideband audio most modern voice models are trained on. Traditional phone lines are narrowband: sampled at 8kHz using the G.711 codec, which only carries frequencies up to roughly 3.4kHz. That cutoff falls right on top of the frequencies that distinguish consonants like s, f, and th, which is part of why numbers, names, and addresses are the hardest things for any voice AI system, cascaded or full-duplex, to get right over a phone line. Academic research on telephone speech recognition consistently shows accuracy drops when audio is downsampled to 8kHz.
That said, this is not strictly an architecture problem. OpenAI's own Realtime API, a speech-to-speech model in the same family as GPT-Live, already supports G.711 8kHz audio natively for phone calls, so full-duplex and speech-to-speech models are not technically locked out of narrowband audio. The real reason cascaded pipelines still dominate telephony is track record: speech-to-text vendors have spent years tuning specifically for noisy, narrowband phone audio, and that specialization is hard to replicate quickly. Cascaded is also easier to debug: every step is inspectable, and a compliance team can insert a content filter into the text stream between the language model and the text-to-speech step in a way that is not possible once a single model generates audio directly.
That is why the newest architecture is not automatically the right one for every deployment. A cold outbound script that has to read a disclosure verbatim for compliance reasons is usually better served by a system where every step can be checked, not a full-duplex model optimized for natural back-and-forth.
How Should a Team Decide Which Architecture Fits a Given Call Type?
Match the architecture to what the call actually needs, not to whichever one launched most recently.
- High-compliance, scripted calls (debt collection, insurance verification, appointment confirmation): cascaded, because every step can be inspected and controlled.
- High-empathy, unscripted calls (support, concierge, anything where the caller's emotional state matters): full-duplex or turn-based, where naturalness drives satisfaction more than precise control.
- High-volume outbound at low cost: cascaded and turn-based platforms remain the cost leaders. xAI's Voice Agent Builder, launched July 1, 2026 at $0.05 per minute, is pushing price down further inside that category rather than in full-duplex.
Frequently Asked Questions
Is full-duplex voice AI the same as speech-to-speech?
They overlap but are not identical. Speech-to-speech usually describes any model that processes audio directly without a text intermediary. Full-duplex specifically means the system listens and generates simultaneously rather than in discrete turns. A model can be speech-to-speech and still be turn-based if it waits for silence before responding.
Is full-duplex voice AI available for building phone agents today?
Not broadly. GPT-Live is rolling out inside ChatGPT Voice, and OpenAI says an API version is planned, with developers and enterprises able to sign up to be notified when it opens, but there is no confirmed date. Most third-party voice AI platforms that agencies build on, including Vapi, Retell, and Bland, still run cascaded pipelines by default for telephony use cases.
Does full-duplex voice AI cost more than cascaded?
Not inherently, but it is newer and less commoditized, so pricing and third-party API access lag behind the mature cascaded ecosystem. As of July 2026, cascaded and turn-based platforms are the price leaders, with Retell at $0.07/min and xAI's new Voice Agent Builder at $0.05/min.
Voxfra sits above the voice AI providers as the routing and reporting layer, which is what lets a team match cascaded, turn-based, or full-duplex to a given client or call type without rebuilding infrastructure every time a provider or architecture changes. See how provider choices get evaluated in practice.