Full-Duplex-Bench v3
Voice agents hear real human speech with five kinds of disfluency and must make chained API calls across four task domains. Scored on Pass@1, response quality, latency and turn-taking.
Code and data were released on May 20th, 2026. The paper evaluates six systems including GPT-Realtime, Gemini Live and Grok.
- Made by
- Guan-Ting Lin, Hung-yi Lee and collaborators (National Taiwan University and others)
- Released
- April 2026
- Kind
- Speech recognition and translation
- Level
- Everyday
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- All public
- Last in a launch post
- May 2026, dropped before it saturated
- Last model added
- May 2026
In the authors’ words
We introduce Full-Duplex-Bench-v3 (FDB-v3), a benchmark for evaluating spoken language models under naturalistic speech conditions and multi-step tool use. Unlike prior work, our dataset consists entirely of real human audio annotated for five disfluency categories, paired with scenarios requiring chained API calls across four task domains. We evaluate six model configurations – GPT-Realtime, Gemini Live 2.5, Gemini Live 3.1, Grok, Ultravox v0.7, and a traditional Cascaded pipeline (Whisper→GPT-4o→TTS) – across accuracy, latency, and turn-taking dimensions. GPT-Realtime leads on Pass@1 (0.600) and interruption avoidance (13.5%); Gemini Live 3.1 achieves the fastest latency (4.25 s) but the lowest turn-take rate (78.0%); and the Cascaded baseline, despite a perfect turn-take rate, incurs the highest latency (10.12 s). Across all systems, self-correction handling and multi-step reasoning under hard scenarios remain the most consistent failure modes.
TML-Interaction-Small leads with 68%
Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.
Source: Labs’ model cards
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 |  TML-Interaction-Small | 68% | May 2026 | Lab model card |
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Labs’ model cards
Labs headline it less often now, last in May 2026
Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.
| Date | Launch post | Score | As the post writes it |
|---|---|---|---|
| May 11, 2026 | TML-Interaction-Small | 68 | FD-bench v3 Pass@1, with background agent, audio plus tools |
Versions
More in Audio and video
- Video-MMMU In use 300 expert-level lecture videos with questions testing perception, comprehension and adaptation of knowledge. Accuracy.
Qwen 3.8 Max 88.7% - LVBench In use Hours-long video understanding with multiple-choice questions on reasoning, grounding and summarization.
Gemini 4 Argon 91.7% - MMVU In use 3,000 expert-annotated questions on videos across 27 disciplines. Multiple-choice and open-ended accuracy. Qwen 3.8 Max 82.4%
- MLVU In use Multi-task long video understanding (minutes to hours) with multiple-choice and generation tasks. Accuracy. Qwen 3.8 Max 90.8%
- FLEURS In use Models transcribe read speech in 102 languages built from FLoRes-101 sentences. Model cards usually report word error rate, so lower scores are better.
Phi-4-multimodal 4.00% - VoiceBench In use Spoken instructions with speaker, environment and content variation. Voice assistant responses scored by GPT judge or rules. Qwen3.5-Omni-Plus 93.1%