You can use full-duplex voice AI in three ways: talk to an app such as ChatGPT Voice, build on a hosted API, or run an open model on your own hardware. The database lists 14 hosted APIs and 18 open models that hold a conversation in full-duplex.
Apps you can talk to
These need no code. The Model column links to the entry in the database.
| Where | Model | What to know |
|---|---|---|
| ChatGPT Voice, in the ChatGPT apps and on ChatGPT.com | GPT-Live | The Live option is the full-duplex one. It can search the web and use connected apps while you talk. |
| Gemini Live, in the Gemini mobile app | Not stated | Google has not said which of its Live models runs it now. It is not in the Gemini web app. |
| Doubao app, call mode | SeedRealtime | ByteDance says the model is fully live in the app's video calls. |
| Grok, in xAI's mobile apps and Tesla vehicles | Grok Voice | xAI says its developer API is built on the voice stack these use. |
| Moshi demo, in the browser | Moshi | A research demo from Kyutai, in English. |
| Sesame research preview, in the browser | Sesame CSM | Talk with the voices Maya and Miles. The demo runs a larger model than the public checkpoint. |
Full-duplex follows the labels in the database, which are explained in Full-duplex vs cascade vs half-duplex. Checked on 2026-11-11.
How to start ChatGPT Voice
- Open the ChatGPT app on iOS or Android, or go to ChatGPT.com.
- Select the Voice icon. It is in the message bar in the app and in the prompt window on the web.
- Allow microphone access if asked. On first use in the app, choose a voice.
- Start speaking. You can talk while ChatGPT is talking.
Settings, then Voice, can show three options. Live is the full-duplex one and runs on GPT-Live-1.1 or the smaller GPT-Live-1.1 Quip, depending on the plan. Advanced and Standard are the older modes that take turns, and OpenAI retires both on December 11, 2026. Live does not support video or screen sharing.
Live has had two Japanese voices, Himari and Rindō, since October 15, 2026. OpenAI publishes no list of supported languages. Usage limits depend on the plan and are listed on OpenAI's help page.
Source: ChatGPT Voice, OpenAI Help Center. Read more: OpenAI
How to start Gemini Live
- Open the Gemini mobile app on Android, iPhone or iPad.
- Tap Live at the bottom of the screen, or swipe left.
- On first use, follow the instructions on screen.
- Start talking.
Google's Live models are in the database as Gemini 3.8 Live and its predecessors, which developers reach through the Live API. In March 2026 Google said Gemini Live in the app ran on Gemini 3.1 Flash Live. Its September post on Gemini 3.8 Live covers the API and does not mention the app.
Source: Talk naturally with Gemini Live, Gemini Apps Help. Read more: Google DeepMind
APIs for developers
A hosted API gives you the model without the app. Your code opens a session, streams the microphone in and plays the reply as it arrives. 14 full-duplex conversation models in the database are offered this way, newest first.
| Model | Japanese | Languages | Price | Released |
|---|---|---|---|---|
 Amazon Nova 2.5 SonicAmazon / AWS | — | — | $3 in, $12 out per 1M audio tokens | 2026-10 |
| Yes | en es fr +17 | $0.08/min per session minute | 2026-10 | |
Phonic-1Phonic | — | 51 | — | 2026-10 |
 Gemini 3.8 LiveGoogle DeepMind | Yes | en ja zh +94 | $3 in, $12 out per 1M audio tokens | 2026-09 |
| Gemini 3.8 Live with Live AvatarGoogle DeepMind | — | 97 | — | 2026-09 |
 Qwen-Omni Realtime (Qwen3.8-Omni-Flash-Realtime)Alibaba Cloud | — | 113 | — | 2026-09 |
 StepAudio 3 RealtimeStepFun, preview | — | — | — | 2026-09 |
 Higgs RealtimeBoson AI | — | 100+ | — | 2026-08 |
| Qwen-Audio Realtime (3.0 / 3.1)Alibaba Cloud | Yes | zh en ja +11 | $6.4 in, $24 out per 1M audio tokens | 2026-08 |
 SeedRealtimeByteDance Seed | — | — | — | 2026-08 |
 GPT-Live 1OpenAI | — | — | $0.05/min per session minute | 2026-07 |
| Yes | en ja zh +52 | $32 in, $64 out per 1M audio tokens | 2026-07 | |
 TML-Interaction-SmallThinking Machines Lab, preview | — | en | — | 2026-05 |
| Amazon Nova 2 SonicAmazon / AWS | — | en es fr +4 | — | 2025-12 |
Prices are the vendors' list prices held in the database. A dash under Japanese means the entry does not list Japanese by name: some vendors give only a count of languages, and OpenAI publishes no list for GPT-Live 1, which has Japanese voices. The Full-duplex label is the database's reading of each vendor's description. OpenAI uses the term only for GPT-Live, and Google's pages do not use it. Analytics compares the hosted APIs on price and benchmark scores.
Open models you can run
18 full-duplex models have public weights. You need your own GPU. J-Moshi and LLM-jp-Moshi-v1, for example, ask for a Linux machine with at least 24 GB of GPU memory. Check the licence before commercial use.
| Model | Japanese | Languages | Licence | Commercial use | Released |
|---|---|---|---|---|---|
GanderOmni-Interaction-Gander | — | — | Apache-2.0 | Yes | 2026-09 |
Realtime-VenusAnt Group (inclusionAI) | — | en zh | Apache-2.0 | Yes | 2026-09 |
| — | en | OpenMDW-1.1 | Yes | 2026-08 | |
DuplexOmniXi'an Jiaotong University / Peking University / Meituan, preview | — | zh en | Apache-2.0 | Yes | 2026-07 |
Lychee-FDHITsz-TMG | — | zh en | Apache-2.0 | Yes | 2026-07 |
BayLing-DuplexICT, Chinese Academy of Sciences | — | — | glm-4-voice License (commercial use after registration) | Yes | 2026-06 |
 Raon-SpeechChat-9BKRAFTON | — | en | CC-BY-NC-4.0 | Non-commercial | 2026-04 |
 Covo-Audio-Chat / Covo-Audio-Chat-FDTencent, preview | — | zh en | Covo-Audio licence (academic use only) | Research only | 2026-03 |
LLM-jp-Moshi-v1NII (LLM-jp) | Yes | ja | Apache-2.0 | Yes | 2026-02 |
 MiniCPM-o 4.5OpenBMB / ModelBest | — | en zh +30 | Apache-2.0 | Yes | 2026-02 |
 PersonaPlexNVIDIA (ADLR) | — | en | MIT (code) / NVIDIA Open Model License (weights) | Yes | 2026-01 |
| Qwen3-OmniAlibaba Cloud | Yes | en zh ja +16 | Apache-2.0 | Yes | 2025-09 |
| Step-Audio 2 miniStepFun | — | zh en | Apache-2.0 | Yes | 2025-08 |
 Kimi-AudioMoonshot AI | — | zh en | MIT (Apache-2.0 for code derived from Qwen2.5-7B) | Yes | 2025-04 |
 Sesame CSMSesame AI | — | en | Apache-2.0 | Yes | 2025-03 |
J-Moshi / J-Moshi-extNagoya University (Higashinaka Lab) | Yes | ja | CC-BY-NC-4.0 | Non-commercial | 2025-01 |
 GLM-4-VoiceZhipu AI (Z.ai) | — | zh en | Apache-2.0 (code) / GLM-4 licence (weights; commercial use after registration) | Unclear | 2024-10 |
 MoshiKyutai | — | en | Apache-2.0 (code) / CC-BY-4.0 (weights) | Yes | 2024-09 |
Japanese
- ChatGPT Voice: the Live option has Japanese voices.
- Hosted APIs that list Japanese: Grok Voice 3.0 (xAI Speech to Speech API), Gemini 3.8 Live, Qwen-Audio Realtime (3.0 / 3.1) and OpenAI Realtime (gpt-realtime-2.1).
- Open models built for Japanese: LLM-jp-Moshi-v1 from NII, under Apache-2.0, and J-Moshi from Nagoya University, for non-commercial use.
- Shizuku-Duplex, announced by Shizuku AI on November 8, 2026, has no public weights, API or demo.
Every model and dataset that lists Japanese is in the table on the Japanese Knowledge page.
How fast it is spreading
Of the 40 full-duplex models in the database, 2 came out in 2024, 9 in 2025 and 29 so far in 2026. OpenAI moved ChatGPT Voice to a full-duplex model in July 2026 and retires its turn-based voice modes in December. The database records releases, not plans, so it cannot say which service is next.
Releases by quarter are charted in Full-duplex vs cascade vs half-duplex.
FAQ
- Which apps and services use full-duplex voice AI?
- ChatGPT Voice is the most widely available: its Live option runs on OpenAI's full-duplex GPT-Live models. ByteDance's Doubao app uses SeedRealtime for video calls, and Kyutai and Sesame run demos in the browser. For developers, 14 hosted APIs and 18 open models in the database are full-duplex.
- How do I use ChatGPT's voice feature?
- Open ChatGPT and select the Voice icon, in the message bar on iOS and Android or in the prompt window on ChatGPT.com. Allow microphone access and start speaking. Under Settings, Voice, the Live option is the full-duplex one.
- Is ChatGPT's voice mode full-duplex?
- The Live option is. It has run on GPT-Live since July 8, 2026, and OpenAI describes GPT-Live as a full-duplex model that listens and speaks at once. The older Advanced and Standard options take turns, and OpenAI retires them on December 11, 2026.
- Can I use full-duplex voice AI in Japanese?
- Yes. ChatGPT's Live option has had Japanese voices since October 2026. 4 hosted APIs in the database list Japanese (Grok Voice 3.0 (xAI Speech to Speech API), Gemini 3.8 Live, Qwen-Audio Realtime (3.0 / 3.1) and OpenAI Realtime (gpt-realtime-2.1)), and two open models are built for it: LLM-jp-Moshi-v1, which allows commercial use, and J-Moshi, which is for research.
- Will other AI services get full-duplex voice?
- The database records releases, not plans, so it cannot say which service is next. 29 of its 40 full-duplex models were released in 2026, and Amazon, xAI, Google DeepMind, Alibaba Cloud and OpenAI all sell hosted APIs that any app can build on.
Updated