No single score ranks full-duplex voice models. A phone agent has to finish the task, a companion has to get the timing right, and the benchmarks for the two are different. Pick the measure that matches your product, then compare models inside one run of one benchmark. For phone agents, GPT-Live 1 with GPT-6 Astra backend has the highest task score among the hosted APIs compared here (79.9% on τ-Voice v2), and two vendor-submitted systems score higher. For companions, no model in the benchmark authors' turn-taking table replies in under half a second and also waits through pauses. Gemini 3.1 Flash Live comes closest. For Japanese, the only side-by-side numbers come from one vendor's own test.
Two rules before comparing
- Compare inside one run. The same model scores differently when a different team runs the same benchmark, so numbers from two sources do not line up.
- Check who ran it. A benchmark's authors, a public leaderboard and a model's own team are three different kinds of evidence. Every table below says which it is.
Guide: How full-duplex voice AI is evaluated
Start from the product
| Product | What to compare | Benchmarks |
|---|---|---|
| Phone agent for support or sales | Whether the task gets done under the company's rules, with noise, accents and tool calls. | τ-bench 2.0 (τ-Voice), EVA-Bench, Full-Duplex-Bench v3 |
| Companion, character or tutor | When it starts to speak, whether it waits through a pause, and what it does when interrupted. | Full-Duplex-Bench v1, Full-Duplex-Bench v1.5, DuplexAct |
| Meetings and group conversation | When to speak among several people, and who is being addressed. | MultiPartyTurnBench, MultiTalkBench, Duplex-MPE, MP-Bench |
| Video agents and avatars | Whether it reads the user's nonverbal cues, and whether its own face and voice respond in time. | VideoFDB, Omni-DuplexEval, OmniInteract |
| Japanese products | Behaviour in Japanese conversation, where listeners give backchannels far more often than in English. | Shizuku AI's own benchmark (run by the vendor), M3-DuplexBench, J-Moshi (subjective) |
| Self-hosted | Licence and hardware first, then scores for models with public weights. | dGSLM, Moshi and Freeze-Omni in the turn-taking tables, and Ultravox v0.7 in the tool-call table |
Phone agents
τ-bench 2.0 (τ-Voice) is a public leaderboard from Sierra Research. Its voice set, τ-Voice v2, counts the share of customer-service calls completed under a policy prompt, in English, Spanish, Portuguese, Hindi, Chinese and Korean. Each row is a voice model together with the backend or thinking mode it ran with, so the score belongs to the pair. The table shows the two hosted APIs in this site's comparison that have a result on it. Sierra ran both. The leaderboard also takes runs that vendors submit themselves, and two of those score higher: Ava Zero from Avartha at 86.0% and Phonic-1 at 83.2%.
| Model | Setup | Result |
|---|---|---|
| GPT-Live 1 | with GPT-6 Astra backend | 79.9% |
| Gemini 3.8 Live | Extended Thinking, high | 68.5% |
EVA-Bench from ServiceNow runs 213 scenarios in three enterprise domains. The table shows one of its measures, faithfulness: how closely the agent keeps to its instructions and policy, on clean audio. ServiceNow ran the tests and publishes them on the benchmark's results page. The ranges of OpenAI Realtime (gpt-realtime-2.1) and Gemini 3.8 Live overlap, so their order is not settled.
| Model | Setup | Result | Range |
|---|---|---|---|
| Grok Voice 3.0 (xAI Speech to Speech API) | — | 66.2% | 62.3 to 70.1 |
| OpenAI Realtime (gpt-realtime-2.1) | — | 49.4% | 45.4 to 53.4 |
| Gemini 3.8 Live | Extended Thinking, high | 42.1% | 38.1 to 46.0 |
Full-Duplex-Bench v3 asks for one to three chained tool calls from real recorded requests with hesitations and self-corrections. The table is the benchmark authors' own run from April 2026. Its GPT-Realtime is gpt-realtime-1.5, an earlier version than the one in the EVA-Bench table above, and its Grok row predates Grok Voice 3.0. The test has 100 recordings, so one point is one recording. GPT-Realtime got the most fully right (60%). Gemini Live 3.1 answered soonest (4.25 s) but gave no answer to 22 of the 100 requests. Higher is better for the first two columns and lower for the other two.
| Model | All tool calls right (Pass@1) | Gave an answer | Time to answer | Cut the user off |
|---|---|---|---|---|
| GPT-Realtime | 60% | 96% | 6.89 s | 14% |
| Gemini Live 3.1 | 54% | 78% | 4.25 s | 19% |
| Gemini Live 2.5 | 49% | 92% | 7.26 s | 14% |
| Cascade baseline Whisper + GPT-4o + OpenAI TTS | 45% | 100% | 10.12 s | 33% |
| Grok | 43% | 94% | 6.65 s | 26% |
| Ultravox v0.7 | 41% | 96% | 8.40 s | 48% |
Guide: Tool calls and search
Companions and characters
Full-Duplex-Bench v1 plays fixed recordings in English and checks turn-taking. The table is the one the benchmark's authors publish: 4 rows from their 2025 paper and 4 they added to the repository README in May 2026. The quickest reply comes first. Reply delay counts only the turns a model took. Higher is better for taking the turn when the user has finished. Lower is better for the other columns. The PersonaPlex row is a pre-release checkpoint, and its numbers are the same as in NVIDIA's own paper. For the released model, that paper reports a reply delay of 0.17 s and 36% of synthetic pauses talked over.
| Model | Reply delay | Talks over a pause, synthetic | Talks over a pause, real | Takes the turn | Delay after an interruption |
|---|---|---|---|---|---|
| PersonaPlex pre-release model | 0.07 s | 58% | 66% | 99% | 0.40 s |
| Moshi | 0.27 s | 99% | 98% | 94% | 0.26 s |
| dGSLM | 0.35 s | 93% | 94% | 98% | 2.53 s |
| Gemini 3.1 Flash Live minimal thinking | 0.57 s | 2% | 15% | 100% | 0.34 s |
| Gemini 3.1 Flash Live high thinking | 0.94 s | 1% | 10% | 99% | 0.65 s |
| Freeze-Omni | 0.95 s | 64% | 48% | 34% | 1.41 s |
| Gemini 2.0 Flash Live | 1.30 s | 26% | 31% | 66% | 1.18 s |
| GPT-Realtime version not stated | 1.47 s | 1% | 12% | 100% | 1.50 s |
dGSLM, Moshi and PersonaPlex (pre-release model) replied in under half a second and talked over 58% to 99% of the synthetic pauses. GPT-Realtime and Gemini 3.1 Flash Live talked over 1% to 2% of the synthetic pauses and 10% to 15% of the pauses cut from real conversations, and took 0.57 to 1.47 seconds to reply. No model in this table did both. Gemini 3.1 Flash Live (minimal thinking) came closest, at 0.57 seconds and 2%.
The quick group are research models with open weights. The patient group are hosted APIs added to the table a year later, so part of the gap may come from when and how each was run. The authors have not yet run GPT-Live 1, Gemini 3.8 Live or Grok Voice 3.0.
The table scores timing only. On the benchmark's 0 to 5 rating of the reply after an interruption, dGSLM scored 0.2 and Moshi scored 0.8, against 3.4 to 4.2 for the others. If cutting in on a user who is still thinking is the worse failure for your product, start from the second group.
Full-Duplex-Bench v1.5 overlaps the model's speech with four kinds of sound: a user interruption, a user backchannel, the user talking to someone else, and background speech. The table shows three of them. It is the authors' run from the 2025 paper, so the models are older. Higher is better for every column except time to stop.
| Model | Responds to an interruption | Time to stop | Keeps going through a backchannel | Keeps going through background speech |
|---|---|---|---|---|
| GPT-4o Realtime 2024-12-17 preview | 78% | 0.23 s | 70% | 4% |
| Freeze-Omni | 72% | 1.42 s | 80% | 25% |
| Moshi | 50% | 1.16 s | 6% | 7% |
| Gemini 2.0 Flash Live | 33% | 2.20 s | 93% | 30% |
| Nova Sonic v1 | 24% | 2.25 s | 98% | 98% |
GPT-4o Realtime (2024-12-17 preview) responded to the most interruptions (78%), but kept going through only 4% of background speech. Nova Sonic (v1) kept going through 98% of backchannels and 98% of background speech, and responded to 24% of interruptions. Moshi's 6% is not a count of stops: the judge could not classify what Moshi did after 92% of backchannels.
DuplexAct checked whether a model gives a backchannel at the moment one is due. Most of its 14 models almost never did. The highest, Raon-SpeechChat, did so 31% of the time, and GPT-Realtime-2.1 0.2%. See Backchannels and Interruption and barge-in.
Group conversation and video
- MultiPartyTurnBench tests two to five speakers in 240 synthetic conversations. Five full-duplex agents usually knew when to start speaking (86% to 96% balanced accuracy). They did no better than chance at telling a real interruption from other speech (47% to 55%). Gemini 3.8 Live stopped for 97.7% of backchannels.
- Duplex-MPE tested eight open-weight systems with three or four people talking. MiniCPM-o 4.5 led on three of its four scores.
- MP-Bench tested 12 voice agents as one participant in a group. The real-time agents scored 22% or lower on following who said what, and near chance on turn-taking.
- VideoFDB is a benchmark for full-duplex video agents from NVIDIA and David AI. A result reaches its leaderboard when a team scores recordings of its agent with the public scorer and sends the videos and scores to NVIDIA. Griffin-Lite from Tavus is the highest-scoring system after the human reference on both tracks. It scored 3.83 out of 5 for generation (human reference 3.92) and 3.73 for perception (human reference 4.20). Its median response took 1,892 ms and 2,232 ms on the two tracks, against 900 ms and 1,400 ms for the human references. Griffin-Lite is open to selected testers only.
Japanese
The Full-Duplex-Bench tables above were measured in English, and the six languages of τ-Voice do not include Japanese. Side-by-side numbers for Japanese are scarce. The fullest set comes from Shizuku AI, which built a Japanese benchmark and ran it on its own model and five others. Nobody outside the company can rerun it yet, because neither the model nor the test set is public.
- Among the four models it classed as full-duplex, Shizuku-Duplex had the best result on five of six turn-taking measures. On two of those five the best result was still low: 17% for staying quiet when the user talks to someone else, and 25% for ignoring background speech. Shizuku-Duplex has not been released.
- GPT-Live 1 had the best content scores of the four: 95% on multi-turn questions, against 79% for Shizuku-Duplex.
- The benchmark classed GPT-Realtime-2.1 and Gemini 3.8 Live as not full-duplex, because both reply after detecting the end of the user's speech. This site's database lists both as full-duplex. They scored 96% and 98% on multi-turn questions and always stopped when interrupted. They kept talking through 0% and 13% of user backchannels.
- J-Moshi-ext and LLM-jp-Moshi-v1, the two open Japanese models in the test, scored 1% on multi-turn questions. Both are prototypes trained on casual conversation. J-Moshi-ext is licensed for non-commercial use only.
M3-DuplexBench from NTT scores turn-taking, pause handling, user backchannels and barge-in in English and Japanese. Its code is under an evaluation licence for non-commercial use.
J-Moshi (subjective) is a human rating on a five-point scale. In the J-Moshi paper, J-Moshi-ext was rated 2.66 for naturalness against 2.44 for a dGSLM baseline. In a separate test that NII ran itself, LLM-jp-Moshi-v1 was rated 3.34 against 2.43 for J-Moshi.
Hosted APIs and open models that list Japanese: Where to use full-duplex voice AI. Read more: Shizuku AI, NII, Nagoya, and J-Moshi
Response time
Vendors' own latency figures measure different things under different conditions, so they do not line up. The table uses one definition, the reply delay of Full-Duplex-Bench v1, as each vendor measured it on its own model. The two OpenAI rows come from one OpenAI run. The Qwen row comes from Alibaba's own paper, so by the first rule the table cannot rank Qwen against the OpenAI rows. For reference, the benchmark's authors measured 0.57 s for Gemini 3.1 Flash Live (minimal thinking), the quickest hosted API in their table.
| Model | Result | Source |
|---|---|---|
| GPT-Live 1 | 0.80 s | OpenAI's own post |
| OpenAI Realtime (gpt-realtime-2.1) | 1.41 s | OpenAI's own post |
| Qwen-Audio Realtime (3.0 / 3.1) | 1.92 s | Alibaba Cloud's own paper |
Analytics charts every run in the database on its own tab and names its source.
Narrowing the list
- Language, price and how each model is offered: Where to use full-duplex voice AI.
- Models with public weights, 27 in the database, 17 of them open to commercial use: Open models you can run.
- Whether a cascade would do: How full-duplex differs from conventional voice AI.
Read more: Why STS needs new benchmarks, Full-Duplex-Bench
FAQ
- Which full-duplex voice model is the best?
- There is no single ranking. Benchmarks measure different things, and the same model scores differently depending on who ran the test. Choose the measure that matches your product, then compare models inside one run of one benchmark.
- Which full-duplex model is best for customer-service calls?
- On τ-Voice v2, Sierra's public leaderboard of customer-service calls, the two highest results in the database were submitted by the vendors themselves: Ava Zero from Avartha at 86.0% and Phonic-1 at 83.2%. Of the hosted APIs compared on this page, which Sierra ran itself, GPT-Live 1 with GPT-6 Astra backend is highest at 79.9%. On EVA-Bench's faithfulness measure, which scores how closely an agent keeps to its instructions, the highest is Grok Voice 3.0 (xAI Speech to Speech API) at 66.2%. Both score a voice model together with the setup it was run in.
- Which full-duplex model is best for a companion or character?
- It depends on which mistake the product can live with. In the table the Full-Duplex-Bench v1 authors publish: dGSLM, Moshi and PersonaPlex (pre-release model) replied in under half a second and talked over 58% to 99% of the synthetic pauses. GPT-Realtime and Gemini 3.1 Flash Live talked over 1% to 2% of the synthetic pauses and 10% to 15% of the pauses cut from real conversations, and took 0.57 to 1.47 seconds to reply. No model in this table did both. Gemini 3.1 Flash Live (minimal thinking) came closest, at 0.57 seconds and 2%.
- Which full-duplex model works best in Japanese?
- The fullest side-by-side numbers come from Shizuku AI's own benchmark of six models, which nobody outside the company can rerun yet. Among the four models it classed as full-duplex, GPT-Live 1 had the best content scores and Shizuku-Duplex had the best result on five of six turn-taking measures. Shizuku-Duplex has not been released. The benchmark classed GPT-Realtime-2.1 and Gemini 3.8 Live as not full-duplex. Those two scored 96% and 98% on multi-turn questions, above GPT-Live 1 at 95%, and kept talking through 0% and 13% of user backchannels. The open Japanese models J-Moshi-ext and LLM-jp-Moshi-v1 scored 1% on multi-turn questions.
- Which full-duplex model responds fastest?
- It depends on the test and on who ran it. In the table the Full-Duplex-Bench v1 authors publish, the lowest reply delay is 0.07 seconds, for a PersonaPlex checkpoint that was never released. The figure matches NVIDIA's own paper, which reports 0.17 seconds for the released model. The quickest hosted API in the authors' table is Gemini 3.1 Flash Live (minimal thinking), at 0.57 seconds. OpenAI ran the same benchmark itself and reported 0.80 seconds for GPT-Live 1. Vendors' own latency figures measure different things and should not be set side by side.
- Which full-duplex models can I run myself?
- 27 full-duplex models in the database have public weights, and 17 of them allow commercial use. They are listed with licences in Where to use full-duplex voice AI.
- Can I compare scores from two different papers?
- Only if both used the same protocol. Papers often change the rules of a benchmark, for example the judge model or how long a model must speak to count as taking the turn. The database keeps each run on its own tab for that reason.
Updated