Full-duplex voice AI, explained

Half duplex vs full duplex

The terms come from telecoms. A walkie-talkie is half duplex: only one side can talk at a time. A phone call is full duplex: both people can speak at once.

Most voice assistants work like walkie-talkies. They detect when you stop talking, then answer. Many let you interrupt, but they still treat the conversation as a strict sequence of turns.

A full-duplex model keeps listening while it speaks. Kyutai's Moshi paper describes this as modelling the user's speech and the model's own speech as two parallel streams, with no explicit turns. Overlaps, interruptions and short interjections then become a normal part of the conversation instead of errors.

Sources: Kyutai on arXiv

The usual pipeline: a cascade

The Moshi paper describes the spoken dialogue systems of its time as pipelines of separate parts. Many voice agents are built from the same four pieces:

• Voice activity detection decides when you have finished speaking.
• Speech-to-text turns your audio into a transcript.
• A text language model writes the reply.
• Text-to-speech reads the reply aloud.

Sources: Kyutai on arXiv

What cascades do well, and what they lose

Cascades are flexible. You can swap any part, use any text model, and keep a transcript for logs and tool calls. Even Kyutai, the lab behind Moshi, builds this kind of system: its Unmute project gives any text model a voice with streaming speech recognition and speech synthesis.

The cost is in what gets lost between the parts. The Moshi paper names three problems. The stages add up to seconds of delay. Information that changes meaning, such as emotion and non-speech sounds, is lost in the text step. And the system depends on clean speaker turns, so overlaps and interruptions are hard to handle.

Sources: Kyutai, Kyutai on arXiv

Speech-to-speech models

A speech-to-speech model takes audio in and produces audio out in one network. Some still use text inside. Moshi predicts the text of its own speech as an inner monologue, which its authors say improves quality. Block says its new Dessa model uses no text at all.

Speech-to-speech does not always mean full duplex. OpenAI's Realtime API is speech-to-speech, but voice activity detection decides when a turn ends. Google describes Gemini 3.8 Live as a native speech-to-speech model, and its Live API docs say users can interrupt the model at any time, but the docs we checked do not call it full duplex.

Sources: Kyutai on arXiv, Block, OpenAI, OpenAI, Google, Google AI for Developers

Turn-taking, barge-in and backchannels

Three behaviours separate a natural voice agent from a stiff one. Turn-taking is knowing when the other person has finished or is only pausing: cutting in too early is rude, and waiting too long feels slow. Barge-in is what happens when the user interrupts, and the agent should stop or adapt, even on a noisy phone line. Backchannels are short sounds, like mm-hm or a laugh, that show the listener is following without taking the turn.

Turn-based systems handle turn-taking and barge-in with rules. OpenAI's Realtime API, for example, can end a turn after a set length of silence, or use a classifier that estimates from your words whether you are done. Backchannels are hard to add this way, because the model only speaks when it holds the turn.

Full-duplex models learn these behaviours from conversation data, and they can be tuned. On 2026-06-10, Kyutai published a version of Moshi post-trained with reinforcement learning to cut badly timed interruptions, respond faster at the end of a turn and time its backchannels better.

Sources: OpenAI, Kyutai, Kyutai on Hugging Face

Latency

People leave short gaps between turns. A 2009 study of ten languages found that speakers everywhere avoid talking over each other and keep the silence between turns short, with only small differences between languages.

Every stage of a cascade adds delay: waiting to be sure you have finished, transcribing, generating text and starting the voice. Speech-to-speech models remove some of those steps. Kyutai reports a theoretical latency of 160 ms for Moshi, and as little as 200 ms in practice on an NVIDIA L4 GPU. OpenAI reports 0.798 seconds of turn-taking latency for GPT-Live-1 in its own tests, compared with 1.41 seconds for GPT-Realtime-2.1.

Vendors measure latency in different ways, so treat their numbers as rough. What matters is the delay your users hear from end to end, including the network and the phone line.

Sources: PNAS, Kyutai on GitHub, OpenAI Developer Community

Why emotion and timing get lost in text

A transcript keeps the words and drops most of the rest. Sarcasm, a sigh, a laugh, a shaky voice and the length of a pause all change what a sentence means, and none of them survive speech-to-text.

Block puts this at the center of its pitch for Dessa. It says text-based voice AI cannot hear how people feel, and calls that a safety problem as well, for example when a caller is in pain. The Moshi paper makes the same point about emotion and non-speech sounds.

Timing is lost too. A cascade sees finished turns, so it cannot react to a laugh in the middle of your sentence or notice that you hesitated.

Sources: Block, Kyutai on arXiv

Known full-duplex models

Full-duplex speech models are still rare. Research on speech-only dialogue goes back at least to 2022, when the dGSLM model generated two-sided conversations, laughter included, from 2,000 hours of phone calls without any text. It was not an assistant you could talk to.

These are the full-duplex models we could verify as of 2026-10-31:

• Moshi (Kyutai): shown on 2024-07-03 and released with open weights on 2024-09-18. English only, built on the Helium text model, weights under CC-BY 4.0.
• PersonaPlex (NVIDIA): released on 2026-01-15. A fine-tune of Moshi that adds control over voice and role. English only, under the NVIDIA Open Model License.
• GPT-Live-1 (OpenAI): generally available in OpenAI's API since 2026-09-10. A closed model billed per minute, which hands reasoning and tools to a separate backend model.
• Dessa (Block): announced on 2026-10-23. Block says it was trained on speech only. It is not publicly available, and Block is giving early access to a small group of AI safety researchers.

Sources: arXiv, Fortune, Kyutai, Kyutai on Hugging Face, NVIDIA on Hugging Face, OpenAI, OpenAI Developer Community, Block

What to look for in a voice API

Demos are easy to make and hard to compare. We are building a hosted API for full-duplex models, with WebRTC and phone support, and we are not affiliated with Block or any model maker. Whichever provider you consider, check these points:

• Duplex mode: does the model listen while it speaks, or wait for a turn to end?
• Latency: the delay your users hear, from the end of their speech to the first audio back, on your real network.
• Interruptions: does it stop quickly, and does it remember what it was saying?
• Connections: WebRTC for browsers and apps, SIP for phone calls, WebSocket for server-side audio.
• Tools and facts: can it call your APIs, or does it hand off to a text model?
• Languages and voices: which are supported, and can you use your own voice?
• License: open weights you can host yourself, or a closed API only?
• Pricing: per minute or per token, and what counts as billable time?
• Data: what audio and transcripts are stored, and for how long?