Comparison
Full-duplex voice models compared
A full-duplex voice model listens while it talks, so it can be interrupted and can react mid-sentence. Only a few exist. This page compares four that we could verify as of 2026-10-31: two open models you can run yourself, one closed API and one announced model you cannot use yet.
Updated
| Compared | DessaBlock | MoshiKyutai | PersonaPlexNVIDIA | GPT-Live-1OpenAI |
|---|---|---|---|---|
| Availability | Announced 2026-10-23. Early access for AI safety researchers only. | Public since 2024-09-18. | Public since 2026-01-15. | Generally available in OpenAI's API since 2026-09-10. |
| Open weights | None published. | Yes. | Yes, as a gated download on Hugging Face. | No. |
| License | None published. | Weights under CC-BY 4.0. Code under MIT and Apache 2.0. | Weights under the NVIDIA Open Model License, which NVIDIA says is ready for commercial use. Code under MIT. | Closed model under OpenAI's API terms. |
| Training data | Human conversation audio only, no text, per Block. | A text model first, then about 7 million hours of audio, phone calls and synthetic dialogue. | Moshi weights, fine-tuned on Fisher English phone calls and synthetic conversations, under 10,000 hours in total. | Not published. |
| Text backbone | None, per Block. | Helium, a 7B text model, plus an inner text stream. | Helium, inherited from Moshi. | Not published. Reasoning goes to a separate backend model. |
| Model size | Not published. | 7B main Transformer. Hugging Face lists the checkpoint at 8B. | 7B per NVIDIA. Hugging Face lists the checkpoint at 8B. | Not published. |
| Latency | No figure. Faster than real time, per Block. | As little as 200 ms on an NVIDIA L4 GPU, per Kyutai. | No figure in plain units. NVIDIA reports FullDuplexBench scores. | 0.798 seconds of turn-taking latency in OpenAI's own tests. |
| Languages | Not published. | English. | English. | Voices in English, Portuguese, German, French, Hindi, Japanese and Korean. |
| Hardware | MacBook Pro M4 Max or better, per Block. | A GPU with about 24 GB of memory, or Apple silicon through MLX. | Tested on an NVIDIA A100 80 GB under Linux. Can offload layers to the CPU. | OpenAI's cloud only. |
| Hosting options | None today. | Host it yourself or use a provider. CC-BY 4.0 requires attribution. | Host it yourself under NVIDIA's license. | OpenAI's Live API, over WebRTC, WebSocket or SIP. |
Our take
To build today with full control, start with Moshi or PersonaPlex. Check the license of the exact checkpoint you use, and plan for English only.
To ship quickly without running GPUs yourself, GPT-Live-1 is the only full-duplex model on this list that its maker sells as a hosted API, with browser and phone connections built in.
Watch Dessa, but do not plan around it until Block publishes a license and access terms. We plan to host it if the license allows, alongside other full-duplex models.
What counts as full duplex
A full-duplex model handles your audio and its own audio at the same time, as two parallel streams. It does not need a separate detector to decide that you have finished. That is what lets it interrupt, be interrupted and make small listener sounds without taking the turn.
Many speech-to-speech services allow interruptions but still work in turns. OpenAI's Realtime API uses voice activity detection to find the end of a turn, and Google's Gemini Live API docs describe interruptions without calling the model full duplex. We left both out of the table.
Sources: Kyutai on arXiv, OpenAI, Google AI for Developers
The open models: Moshi and PersonaPlex
Moshi comes from Kyutai, an open-science AI lab in Paris, which describes it as the first real-time full-duplex spoken language model. Kyutai released the weights and code on 2024-09-18, a day after its technical report appeared on arXiv. Moshi runs on a single GPU or on a Mac through MLX, and it speaks English only.
PersonaPlex is NVIDIA's fine-tune of Moshi, released on 2026-01-15. It adds a voice prompt and a text prompt, so you can set how it sounds and what role it plays, for example a customer service agent. It is also English only, and NVIDIA tested it on data-center GPUs.
Their licenses differ. Moshi's weights are under CC-BY 4.0, which allows commercial use with attribution, although Kyutai's model card calls the model research-only. PersonaPlex is under the NVIDIA Open Model License, and NVIDIA says it is ready for commercial use. Kyutai's newer reinforcement-learning version of Moshi, which improves turn-taking and backchannels, is non-commercial.
Sources: Kyutai, Fortune, Kyutai on arXiv, Kyutai, Kyutai on GitHub, Kyutai on Hugging Face, NVIDIA on Hugging Face, NVIDIA on GitHub, Kyutai on Hugging Face
The closed API: GPT-Live-1
OpenAI's GPT-Live-1 is the only model in the table that its maker sells as a hosted service. It is full duplex, connects over WebRTC, WebSocket or SIP, and costs $0.05 per minute of voice session, billed per second.
It does not do everything itself. Reasoning, search and tools go to a backend model that you choose and pay for separately. OpenAI reports 0.798 seconds of turn-taking latency in its own tests, and its docs list voices in seven languages.
Sources: OpenAI, OpenAI Developer Community, OpenAI, OpenAI
The newcomer: Dessa
Block announced Dessa on 2026-10-23. According to Block, it was trained from scratch on human conversation only, with no text and no pretrained model, on one NVIDIA GB200 NVL72, and it runs faster than real time on a MacBook Pro M4 Max or better.
If that holds up, Dessa is built differently from the rest of the table. Moshi and PersonaPlex start from a text model, and OpenAI has not said how GPT-Live-1 was trained. But Block has released no weights, API, license or paper, so Dessa cannot be tested yet.
Sources: Block, Kyutai on arXiv, NVIDIA on GitHub
A model is not a voice service
Only GPT-Live-1 arrives as a service, with WebRTC for browsers, WebSocket for servers and SIP for phone calls. The open models give you weights and inference code. Turning them into something callers can reach, with phone numbers, scaling, monitoring and recordings you control, is work you do yourself or hand to a host.
That is also where licenses matter most. A host can only offer a model whose license allows it, which is why no outside host can offer Dessa until Block publishes terms.
Sources: OpenAI, OpenAI, Kyutai on GitHub, NVIDIA on GitHub, Block
Other models people ask about
Google announced Gemini 3.8 Live on 2026-09-15 as a native speech-to-speech model for its Live API, and says it covers more than 90 languages. The Live API docs say users can interrupt the model at any time, but they do not describe it as full duplex, so it is not in the table.
Sesame's open CSM-1B, released on 2025-03-13 under Apache 2.0, is a speech generator rather than a dialogue model. Its README says to pair it with a separate language model for the text, which makes it part of a cascade.
Speech-only dialogue research is older than Dessa. The 2022 dGSLM model learned to generate two-sided conversations, laughter included, from 2,000 hours of phone calls without text, though it was not an assistant you could talk to.
Sources: Google, Google AI for Developers, Sesame on GitHub, arXiv
Full-duplex voice API
Build on full duplex with our hosted API.
Pick a plan, and get Dessa news in the same step if you want it.
FAQ
Questions
Which full-duplex voice model is best?
It depends on what you need. GPT-Live-1 is the quickest to ship, Moshi and PersonaPlex are open, and Dessa is not available. No independent benchmark covers all four.
Are there open-source full-duplex models?
Yes. Kyutai's Moshi has open weights under CC-BY 4.0 and open code. NVIDIA's PersonaPlex, built on Moshi, has open weights under the NVIDIA Open Model License and code under MIT.
Is Gemini Live full duplex?
Google describes Gemini 3.8 Live as a native speech-to-speech model that users can interrupt at any time. The Google docs we checked do not call it full duplex.
Do any of these speak languages other than English?
Moshi and PersonaPlex are English only. OpenAI lists GPT-Live-1 voices in seven languages. Block has not said which languages Dessa supports.
Sources
- Moshi: a speech-text foundation model for real-time dialogue, Kyutai on arXiv, 2024-09-17
- Voice activity detection (VAD), OpenAI
- Gemini Live API overview, Google AI for Developers
- kyutai: open-science AI lab, Kyutai
- French billionaire Xavier Niel is building a ChatGPT competitor with a 'thick French accent', Fortune, 2024-07-04
- Moshi open-source release: run Moshi locally!, Kyutai, 2024-09-18
- kyutai-labs/moshi, Kyutai on GitHub
- kyutai/moshiko-pytorch-bf16, Kyutai on Hugging Face
- nvidia/personaplex-7b-v1, NVIDIA on Hugging Face, 2026-01-15
- NVIDIA/personaplex, NVIDIA on GitHub
- kyutai/moshika-rl-seamless, Kyutai on Hugging Face
- GPT-Live 1 Model, OpenAI
- Introducing GPT-Live-1 in the API, OpenAI Developer Community, 2026-09-10
- Telephony and SIP, OpenAI
- Managing GPT-Live sessions, OpenAI
- Dessa - the Conversation Model trained only on speech, Block
- Build real-time voice applications with Gemini 3.8 Live and 3.5 Transcribe, Google, 2026-09-15
- SesameAILabs/csm: A Conversational Speech Generation Model, Sesame on GitHub
- Generative Spoken Dialogue Language Modeling, arXiv, 2022-03-30
Independent site. Not affiliated with or endorsed by Block, Inc.