File size: 58,135 Bytes
e5187e3
37daf11
 
 
 
e5187e3
 
 
496ab4f
37daf11
e5187e3
 
 
 
 
 
496ab4f
37daf11
 
 
 
 
496ab4f
37daf11
496ab4f
 
 
 
 
 
e5187e3
37daf11
e5187e3
 
 
496ab4f
 
 
37daf11
 
e5187e3
496ab4f
37daf11
496ab4f
e5187e3
 
 
37daf11
e5187e3
496ab4f
e5187e3
 
37daf11
496ab4f
e5187e3
 
37daf11
e5187e3
37daf11
496ab4f
e5187e3
496ab4f
e5187e3
 
37daf11
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
{"slug":"turnbench","name":"TurnBench","maintainer":"Sesame AI Labs / Mundo AI / CMU / oto","tagline":"Scores end-of-turn and interruption detection on 30 hours of hand-labelled dyadic conversation, with conversation type as a controlled variable.","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex"],"metrics":["end-of-turn recall","interruption recall","false-positive rate","median detection latency"],"models_scored":3,"updated":"2026-11","url":"https://turnbench.sesame.com/","paper_url":"https://arxiv.org/abs/2608.25218","leaderboard_url":"https://turnbench.sesame.com/","page_url":"https://www.fullduplex.ai/benchmarks/turnbench"}
{"slug":"fdb-v1","name":"Full-Duplex-Bench v1","maintainer":"NTU / UC Berkeley / UW / MIT","tagline":"Original Full-Duplex-Bench — automatic scoring of the four canonical turn-taking events: when-to-speak, back-channel, interruption success, pause handling.","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex"],"metrics":["turn-taking latency","back-channel timing","interruption success","pause handling"],"models_scored":14,"updated":"2025-03","url":"https://github.com/DanielLin94144/Full-Duplex-Bench","paper_url":"https://arxiv.org/abs/2503.04721","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/fdb-v1"}
{"slug":"fdb-v15","name":"Full-Duplex-Bench v1.5","maintainer":"NTU / UC Berkeley / UW / MIT","tagline":"Mid-2025 refresh of FDB v1 — adds the simultaneous-speech / overlap axis (user interruption, listener back-channel, side conversation, ambient speech).","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex"],"metrics":["overlap duration","mutual-silence rate","interruption recovery"],"models_scored":6,"updated":"2025-08","url":"https://github.com/DanielLin94144/Full-Duplex-Bench","paper_url":"https://arxiv.org/abs/2507.23159","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/fdb-v15"}
{"slug":"fdb-v2","name":"Full-Duplex-Bench v2","maintainer":"NTU / UC Berkeley / UW / MIT","tagline":"Late-2025 WebRTC-based live harness — first FDB release that runs dynamic evaluation against streaming endpoints, plus multi-turn stimuli.","full_duplex":true,"tier":"native","setting":"live","focus":["full-duplex","speech-lm"],"metrics":["streaming latency","multi-turn coherence","overlap handling"],"models_scored":2,"updated":"2026-02","url":"https://github.com/DanielLin94144/Full-Duplex-Bench","paper_url":"https://arxiv.org/abs/2510.07838","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/fdb-v2"}
{"slug":"fdb-v3","name":"Full-Duplex-Bench v3","maintainer":"NTU (NVIDIA advisory)","tagline":"Multi-step tool calling for voice agents, driven by real human speech with natural disfluencies.","full_duplex":true,"tier":"native","setting":"live","focus":["full-duplex","vertical-agent"],"metrics":["Pass@1","tool-call F1","response latency","turn-take rate"],"models_scored":6,"updated":"2026-05","url":"https://github.com/DanielLin94144/Full-Duplex-Bench","paper_url":"https://arxiv.org/abs/2604.04847","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/fdb-v3"}
{"slug":"videofdb","name":"VideoFDB","maintainer":"NVIDIA / David AI","tagline":"Full-duplex benchmark for video calls: does the agent read the user's nonverbal cues, and do its own voice and face respond the way a person's would?","full_duplex":true,"tier":"native","setting":"live","focus":["full-duplex"],"metrics":["perception rubric (0 to 5)","generation rubric (0 to 5)","TOR-Alignment","median response latency"],"models_scored":4,"updated":"2026-10","url":"https://research.nvidia.com/labs/amri/projects/video-fdb/","paper_url":"https://arxiv.org/abs/2605.30256","leaderboard_url":"https://research.nvidia.com/labs/amri/projects/video-fdb/leaderboard.html","page_url":"https://www.fullduplex.ai/benchmarks/videofdb"}
{"slug":"duplexact","name":"DuplexAct","maintainer":"Dong et al.","tagline":"Scores whether full-duplex speech models take the right interaction action when a conversation calls for it, and whether they follow through.","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex"],"metrics":["immediate execution success","follow-through success"],"models_scored":8,"updated":"2026-10","url":"https://github.com/JiahengDong/DuplexAct","paper_url":"https://arxiv.org/abs/2610.17738","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/duplexact"}
{"slug":"multipartyturnbench","name":"MultiPartyTurnBench","maintainer":"Mathur and Manocha","tagline":"Floor control with two to five speakers: can a detector or full-duplex agent tell a real turn change from a hold, a backchannel, or speech that does not take the floor?","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex"],"metrics":["end-of-turn recall at fixed FPR","interruption recall at fixed FPR","signed latency","selective interruption balanced accuracy"],"models_scored":2,"updated":"2026-10","url":"https://multipartyturnbench.github.io/","paper_url":"https://arxiv.org/abs/2610.20172","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/multipartyturnbench"}
{"slug":"multitalkbench","name":"MultiTalkBench","maintainer":"Wang et al. (MultiTalk)","tagline":"Long, multi-party, English and Chinese full-duplex dialogue, with conversations that average 32.6 minutes.","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex"],"metrics":["long-range entity tracking","topic coherence","addressee selection","turn-taking"],"models_scored":0,"updated":"2026-09","url":"https://huggingface.co/datasets/MultiTalk/MultiTalkBench","paper_url":"https://arxiv.org/abs/2609.36903","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/multitalkbench"}
{"slug":"m3-duplexbench","name":"M3-DuplexBench","maintainer":"NTT (Fukuda et al.)","tagline":"English and Japanese full-duplex benchmark that scores interaction timing and response content together, in casual chat and multi-turn QA.","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex","speech-lm"],"metrics":["turn-taking rate and latency","pause handling","backchannel and barge-in response","response content"],"models_scored":2,"updated":"2026-10","url":"https://github.com/nttcslab-sp/m3-duplexbench","paper_url":"https://arxiv.org/abs/2607.29125","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/m3-duplexbench"}
{"slug":"duplex-mpe","name":"Duplex-MPE","maintainer":"Ma et al.","tagline":"Multi-party full-duplex test: with three or four people talking, should the assistant answer, stay silent or stop?","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex"],"metrics":["response initiation","answer accuracy","silence preservation","stopping when resolved"],"models_scored":0,"updated":"2026-11","url":"https://step-out.github.io/Duplex-MPE-Page/","paper_url":"https://arxiv.org/abs/2609.31948","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/duplex-mpe"}
{"slug":"duplexledger","name":"DuplexLedger","maintainer":"Li et al.","tagline":"Does a full-duplex agent remember the right things? Write, update, ignore and retain decisions during live interaction.","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex","speech-lm"],"metrics":["latest-value accuracy","final-state fact coverage","stale assertions","distractor adoption"],"models_scored":0,"updated":"2026-10","url":"https://github.com/lyx-researcher/DuplexLedger","paper_url":"https://arxiv.org/abs/2610.27581","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/duplexledger"}
{"slug":"interactionbench","name":"InteractionBench","maintainer":"Avera Labs (He et al.)","tagline":"Timing, spoken-task accuracy, whisper and volume, and groundedness over 20+ turns, measured on the same recorded user audio for every system.","full_duplex":true,"tier":"native","setting":"live","focus":["full-duplex","speech-lm"],"metrics":["conversational timing","spoken-task accuracy","whisper and volume control","groundedness"],"models_scored":0,"updated":"2026-11","url":"https://avera-labs.github.io/InteractionBench/","paper_url":"https://arxiv.org/abs/2611.02430","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/interactionbench"}
{"slug":"echo-turn-taking","name":"ECHO","maintainer":"Zhao et al.","tagline":"Paired Chinese turn-taking test: the same overlapping words must sometimes take the floor and sometimes not, depending on context.","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex"],"metrics":["pair accuracy","yield accuracy","keep accuracy","off-talk false yields"],"models_scored":0,"updated":"2026-09","url":"https://huggingface.co/datasets/shuofeng123/ECHO","paper_url":"https://arxiv.org/abs/2609.17360","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/echo-turn-taking"}
{"slug":"mp-bench","name":"MP-Bench","maintainer":"Shih et al.","tagline":"The voice agent is one participant in a spoken group conversation: does it follow who said what, and does it take its turn?","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex","speech-lm"],"metrics":["turn-taking accuracy","response appropriateness","comprehension accuracy","latency"],"models_scored":0,"updated":"2026-10","url":"https://github.com/atosystem/MP-Bench","paper_url":"https://arxiv.org/abs/2609.13076","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/mp-bench"}
{"slug":"dsb-ifeval","name":"DuplexSpeechBench-IFEval","maintainer":"Mathur and Manocha","tagline":"Does a full-duplex agent infer from its persona when to speak, backchannel, interrupt or yield, and not only from explicit rules?","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex","instruction-following"],"metrics":["Instruction Adherence Score","Persona Adherence Score"],"models_scored":0,"updated":"2026-09","url":"https://dsb-ifeval.github.io/","paper_url":"https://arxiv.org/abs/2609.03423","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/dsb-ifeval"}
{"slug":"talking-turns","name":"Talking Turns","maintainer":"Apple / CMU","tagline":"A supervised turn-taking judge trained on human-human conversations to score audio foundation models.","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex","speech-lm"],"metrics":["turn-event F1","backchannel rate","interruption rate"],"models_scored":3,"updated":"2025-04","url":"https://machinelearning.apple.com/research/talking-turns","paper_url":"https://arxiv.org/abs/2503.01174","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/talking-turns"}
{"slug":"voicebench","name":"VoiceBench","maintainer":"NUS","tagline":"Evaluates LLM-based voice assistants across knowledge, reasoning, safety, and instruction-following — judged by a GPT-4-class model.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","instruction-following"],"metrics":["overall","QA","reasoning (BBH)","IFEval","safety"],"models_scored":24,"updated":"2025-06","url":"https://github.com/MatthewCYM/VoiceBench","paper_url":"https://arxiv.org/abs/2410.17196","leaderboard_url":"https://matthewcym.github.io/VoiceBench/","page_url":"https://www.fullduplex.ai/benchmarks/voicebench"}
{"slug":"vocalbench","name":"VocalBench","maintainer":"SJTU (OmniAgent)","tagline":"A four-dimensional stress test for speech-interaction models — semantic, acoustic, conversational, and robust.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","audio-understanding"],"metrics":["ACC","LLM-judge","UTMOS","WER","EER"],"models_scored":1,"updated":"2026-01","url":"https://github.com/SJTU-OmniAgent/VocalBench","paper_url":"https://arxiv.org/abs/2505.15727","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/vocalbench"}
{"slug":"uro-bench","name":"URO-Bench","maintainer":"Ruiqi Yan et al.","tagline":"The first end-to-end S2S benchmark with multilingual, multi-turn, and paralinguistic coverage.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","instruction-following","emotion"],"metrics":["U/R/O composite","paralinguistic score"],"models_scored":5,"updated":"2025-08","url":"https://github.com/Ruiqi-Yan/URO-Bench","paper_url":"https://arxiv.org/abs/2502.17810","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/uro-bench"}
{"slug":"air-bench","name":"AIR-Bench","maintainer":"Alibaba DAMO (OFA-Sys)","tagline":"A two-tier stress test for audio-LLMs covering speech, natural sound, and music — MCQ plus chat.","full_duplex":false,"tier":"native","setting":"lab","focus":["audio-understanding","instruction-following"],"metrics":["MCQ acc","GPT-4 judge score"],"models_scored":4,"updated":"2024-12","url":"https://github.com/OFA-Sys/AIR-Bench","paper_url":"https://arxiv.org/abs/2402.07729","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/air-bench"}
{"slug":"big-bench-audio","name":"Big Bench Audio","maintainer":"Artificial Analysis","tagline":"1 000 audio-question reasoning benchmark — the dataset behind the Artificial Analysis S2S speech-reasoning leaderboard.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","instruction-following"],"metrics":["accuracy","TTFA (s)"],"models_scored":12,"updated":"2026-05","url":"https://huggingface.co/datasets/ArtificialAnalysis/big_bench_audio","paper_url":null,"leaderboard_url":"https://artificialanalysis.ai/speech-to-speech","page_url":"https://www.fullduplex.ai/benchmarks/big-bench-audio"}
{"slug":"aa-speech-to-speech-index","name":"Artificial Analysis Speech to Speech Index","maintainer":"Artificial Analysis","tagline":"Composite score for native speech-to-speech models: speech reasoning, agentic task completion, human preference and task success, weighted equally.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","vertical-agent"],"metrics":["index score","Big Bench Audio accuracy","τ-Voice pass rate","Arena Score","task success rate"],"models_scored":0,"updated":"2026-08","url":"https://artificialanalysis.ai/methodology/speech-to-speech-benchmarking","paper_url":null,"leaderboard_url":"https://artificialanalysis.ai/speech-to-speech","page_url":"https://www.fullduplex.ai/benchmarks/aa-speech-to-speech-index"}
{"slug":"mmar","name":"MMAR","maintainer":"SJTU / MBZUAI · NeurIPS 2025","tagline":"1 000 human-curated deep-reasoning audio QA items spanning speech, sound, music, and their mix — four hierarchical reasoning layers.","full_duplex":false,"tier":"component","setting":"lab","focus":["audio-understanding","speech-lm"],"metrics":["overall accuracy","per-layer accuracy","rubric score"],"models_scored":5,"updated":"2026-02","url":"https://github.com/ddlBoJack/MMAR","paper_url":"https://arxiv.org/abs/2505.13032","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/mmar"}
{"slug":"ifeval-audio","name":"IFEval-Audio","maintainer":"AudioLLMs (I2R Singapore)","tagline":"Format-constrained spoken instructions, verified programmatically — an audio port of text IFEval.","full_duplex":false,"tier":"native","setting":"lab","focus":["instruction-following","speech-lm"],"metrics":["IFR","SCR","OSR"],"models_scored":0,"updated":"2025-05","url":"https://github.com/AudioLLMs/AudioBench/tree/main/IFEval-Audio","paper_url":"https://arxiv.org/abs/2505.16774","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/ifeval-audio"}
{"slug":"scale-voice-showdown","name":"Scale Voice Showdown","maintainer":"Scale AI","tagline":"First in-the-wild preference arena for voice AI, launched March 2026 with 11 frontier models, 60+ languages and real user speech. The public leaderboard is offline as of October 2026.","full_duplex":false,"tier":"native","setting":"arena","focus":["speech-lm","tts"],"metrics":["preference win-rate","per-language score"],"models_scored":1,"updated":"2026-03","url":"https://labs.scale.com/","paper_url":null,"leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/scale-voice-showdown"}
{"slug":"aa-speech-arena","name":"Artificial Analysis Speech Arena","maintainer":"Artificial Analysis","tagline":"Blind user votes turn TTS models into a live Elo leaderboard — the de-facto preference number for TTS. Despite the 'Speech Arena' label, this is a TTS-only (text → speech) arena, not an STS ranking.","full_duplex":false,"tier":"adjacent","setting":"arena","focus":["tts"],"metrics":["Elo"],"models_scored":0,"updated":"2026-10","url":"https://artificialanalysis.ai/text-to-speech/arena","paper_url":null,"leaderboard_url":"https://artificialanalysis.ai/text-to-speech/leaderboard","page_url":"https://www.fullduplex.ai/benchmarks/aa-speech-arena"}
{"slug":"aa-speech-agent-arena","name":"Artificial Analysis Speech Agent Arena","maintainer":"Artificial Analysis","tagline":"Human participants complete the same task with two hidden speech-to-speech systems over live calls, then vote for the conversation they preferred.","full_duplex":false,"tier":"native","setting":"arena","focus":["speech-lm","vertical-agent"],"metrics":["preference Elo","agentic Elo","non-agentic Elo","task success rate"],"models_scored":0,"updated":"2026-10","url":"https://artificialanalysis.ai/speech-to-speech/arena/overview","paper_url":null,"leaderboard_url":"https://artificialanalysis.ai/speech-to-speech/arena","page_url":"https://www.fullduplex.ai/benchmarks/aa-speech-agent-arena"}
{"slug":"rw-voice-eq","name":"Real World Voice EQ Bench","maintainer":"Hume AI","tagline":"Human-rated leaderboards across TTS, speech-to-speech, speech understanding and ASR, built to test whether systems use what is in the audio and not only the words.","full_duplex":false,"tier":"component","setting":"arena","focus":["tts","speech-lm","asr","emotion"],"metrics":["rater score (1 to 5) per factor","WER (ASR board)"],"models_scored":0,"updated":"2026-07","url":"https://huggingface.co/spaces/HumeAI/rw-voice-eq","paper_url":"https://arxiv.org/abs/2607.14846","leaderboard_url":"https://huggingface.co/spaces/HumeAI/rw-voice-eq","page_url":"https://www.fullduplex.ai/benchmarks/rw-voice-eq"}
{"slug":"mtalk-bench","name":"MTalk-Bench","maintainer":"FreedomIntelligence / CUHK-SZ","tagline":"First multi-turn S2S benchmark with both arena-style (pairwise) and rubric-based (absolute) protocols for dialogue evaluation.","full_duplex":false,"tier":"native","setting":"arena","focus":["speech-lm","instruction-following","emotion"],"metrics":["arena win-rate","rubric score","per-dimension score"],"models_scored":0,"updated":"2025-09","url":"https://github.com/FreedomIntelligence/MTalk-Bench","paper_url":"https://arxiv.org/abs/2508.18240","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/mtalk-bench"}
{"slug":"tts-arena","name":"TTS Arena","maintainer":"Hugging Face","tagline":"The original blind A/B voting space for TTS, feeding an Elo ladder on Hugging Face.","full_duplex":false,"tier":"adjacent","setting":"arena","focus":["tts"],"metrics":["Elo"],"models_scored":0,"updated":"2026-03","url":"https://huggingface.co/spaces/TTS-AGI/TTS-Arena-V2","paper_url":null,"leaderboard_url":"https://huggingface.co/spaces/TTS-AGI/TTS-Arena-V2","page_url":"https://www.fullduplex.ai/benchmarks/tts-arena"}
{"slug":"tau3-bench","name":"τ-bench 2.0 (τ-Voice)","maintainer":"Sierra Research","tagline":"The first agent benchmark with a live, full-duplex voice track, now in six languages with version 2.0.","full_duplex":true,"tier":"native","setting":"vertical","focus":["vertical-agent","full-duplex","speech-lm"],"metrics":["pass^k","tool-use success","voice task success","per-language task completion"],"models_scored":5,"updated":"2026-10","url":"https://sierra.ai/blog/bench-advancing-agent-benchmarking-to-knowledge-and-voice","paper_url":"https://arxiv.org/abs/2603.13686","leaderboard_url":"https://www.taubench.com","page_url":"https://www.fullduplex.ai/benchmarks/tau3-bench"}
{"slug":"eva-bench","name":"EVA-Bench","maintainer":"ServiceNow","tagline":"End-to-end voice-agent evaluation with bot-to-bot calls, separate accuracy and experience scores, and accent and noise perturbations.","full_duplex":false,"tier":"native","setting":"vertical","focus":["vertical-agent","speech-lm"],"metrics":["EVA-A (accuracy)","EVA-X (experience)","pass@k and pass^k","accent and noise robustness"],"models_scored":1,"updated":"2026-10","url":"https://servicenow.github.io/eva/","paper_url":"https://arxiv.org/abs/2605.13841","leaderboard_url":"https://servicenow.github.io/eva/","page_url":"https://www.fullduplex.ai/benchmarks/eva-bench"}
{"slug":"vamos-bench","name":"VAmoS Bench","maintainer":"Veris AI","tagline":"Simulated phone calls graded on what the voice agent said and on what it actually changed in the backend.","full_duplex":false,"tier":"native","setting":"vertical","focus":["vertical-agent"],"metrics":["task completion","latency","barge-ins per call","cost per call"],"models_scored":0,"updated":"2026-10","url":"https://veris.ai/leaderboard/vamos","paper_url":"https://arxiv.org/abs/2609.38512","leaderboard_url":"https://veris.ai/leaderboard/vamos-pro","page_url":"https://www.fullduplex.ai/benchmarks/vamos-bench"}
{"slug":"apex-voice","name":"APEX-Voice","maintainer":"Mathur and Manocha","tagline":"Can a full-duplex voice agent finish professional work, such as form completion or negotiation, while the user interrupts and corrects?","full_duplex":true,"tier":"native","setting":"vertical","focus":["vertical-agent","full-duplex"],"metrics":["Pass@1","Reliable@3","artifact field accuracy"],"models_scored":0,"updated":"2026-09","url":"https://apex-voice.github.io/","paper_url":"https://arxiv.org/abs/2609.34973","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/apex-voice"}
{"slug":"voiceagenteval","name":"VoiceAgentEval","maintainer":"Meituan / Xbench / Agora","tagline":"A dual-dimensional benchmark for expert-level outbound-calling agents — interaction fluency vs task flow compliance.","full_duplex":false,"tier":"native","setting":"vertical","focus":["vertical-agent","speech-lm"],"metrics":["GIC","TFC","task success"],"models_scored":0,"updated":"2025-10","url":"https://www.shengwang.cn/news/blogdetail/VoiceAgentEval/","paper_url":"https://arxiv.org/abs/2510.21244","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/voiceagenteval"}
{"slug":"healthbench","name":"HealthBench","maintainer":"OpenAI","tagline":"5,000 multi-turn health conversations × 48k physician-written rubric criteria — text-only, listed as a transferable foundation rubric rather than a voice benchmark.","full_duplex":false,"tier":"adjacent","setting":"vertical","focus":["vertical-agent","instruction-following"],"metrics":["rubric score","per-theme breakdown"],"models_scored":0,"updated":"2025-05","url":"https://huggingface.co/datasets/openai/healthbench","paper_url":"https://arxiv.org/abs/2505.08775","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/healthbench"}
{"slug":"seamless-expressive","name":"SeamlessExpressive (mExpresso)","maintainer":"Meta AI","tagline":"Expressive speech-to-speech translation judged on translation, prosody, and vocal-style preservation.","full_duplex":false,"tier":"component","setting":"lab","focus":["s2st"],"metrics":["ASR-BLEU","AutoPCP","Vocal Style Sim"],"models_scored":0,"updated":"2023-12","url":"https://github.com/facebookresearch/seamless_communication","paper_url":"https://arxiv.org/abs/2312.05187","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/seamless-expressive"}
{"slug":"covost2","name":"CoVoST-2 BLEU","maintainer":"Meta AI","tagline":"21 language pairs of speech-to-text translation, long the S2T standard.","full_duplex":false,"tier":"component","setting":"lab","focus":["s2st","asr"],"metrics":["BLEU","chrF"],"models_scored":5,"updated":"2020-07","url":"https://github.com/facebookresearch/covost","paper_url":"https://arxiv.org/abs/2007.10310","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/covost2"}
{"slug":"seed-tts-eval","name":"Seed-TTS Eval","maintainer":"ByteDance","tagline":"The de-facto evaluation set for zero-shot voice cloning — WER, speaker similarity, and CMOS in one recipe.","full_duplex":false,"tier":"component","setting":"lab","focus":["tts"],"metrics":["WER","SECS","CMOS"],"models_scored":3,"updated":"2024-11","url":"https://bytedancespeech.github.io/seedtts_tech_report/","paper_url":"https://arxiv.org/abs/2406.02430","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/seed-tts-eval"}
{"slug":"nv-bench","name":"NV-Bench","maintainer":"Ni et al.","tagline":"Standardised test for generating laughs, sighs, fillers, and back-channels in TTS.","full_duplex":false,"tier":"component","setting":"lab","focus":["tts","emotion"],"metrics":["PCER","FAD","SIM","DNSMOS"],"models_scored":0,"updated":"2026-03","url":"https://arxiv.org/abs/2603.15352","paper_url":"https://arxiv.org/abs/2603.15352","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/nv-bench"}
{"slug":"superb","name":"SUPERB","maintainer":"NTU et al. (s3prl)","tagline":"The classic: general-purpose speech representations scored on ten frozen-encoder downstream tasks.","full_duplex":false,"tier":"peripheral","setting":"lab","focus":["speech-lm","asr","audio-understanding"],"metrics":["composite score","per-task scores"],"models_scored":0,"updated":"2024-06","url":"https://github.com/s3prl/s3prl","paper_url":"https://arxiv.org/abs/2105.01051","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/superb"}
{"slug":"librispeech","name":"LibriSpeech WER","maintainer":"OpenSLR","tagline":"The coordinate system of English ASR — test-clean and test-other WER.","full_duplex":false,"tier":"peripheral","setting":"lab","focus":["asr"],"metrics":["WER (test-clean)","WER (test-other)"],"models_scored":13,"updated":"2024-01","url":"https://www.openslr.org/12","paper_url":null,"leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/librispeech"}
{"slug":"meld","name":"MELD / IEMOCAP","maintainer":"Academic","tagline":"The classic emotion-recognition pair, evaluated with full dialogue context.","full_duplex":false,"tier":"peripheral","setting":"lab","focus":["emotion","audio-understanding"],"metrics":["Weighted F1","Accuracy"],"models_scored":0,"updated":"2023-06","url":"https://affective-meld.github.io","paper_url":null,"leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/meld"}
{"slug":"sid-bench","name":"SID-Bench","maintainer":"Xia et al.","tagline":"Semantic-aware interruption detection, built from real human dialogues and scored with a time penalty for both false alarms and late responses.","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex"],"metrics":["Average Penalty Time (APT)","false alarms","response delay"],"models_scored":0,"updated":"2026-03","url":"https://github.com/xkx-hub/SID-bench","paper_url":"https://arxiv.org/abs/2603.24144","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/sid-bench"}
{"slug":"fd-bench","name":"FD-Bench","maintainer":"Peng et al.","tagline":"A pipeline of LLM, TTS and ASR that simulates users and tests how full-duplex systems handle interruptions, delays and noise.","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex"],"metrics":["interruption response","delay","robustness under noise"],"models_scored":0,"updated":"2025-07","url":"https://github.com/pengyizhou/FD-Bench","paper_url":"https://arxiv.org/abs/2507.19040","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/fd-bench"}
{"slug":"mtr-duplex-bench","name":"MTR-DuplexBench","maintainer":"Zhang et al.","tagline":"Multi-round benchmark that splits continuous full-duplex dialogue into turns and scores each round.","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex","speech-lm"],"metrics":["conversational features","dialogue quality","instruction following","safety"],"models_scored":0,"updated":"2026-04","url":"https://github.com/ZhangHe0918/MTR-DuplexBench","paper_url":"https://arxiv.org/abs/2511.10262","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/mtr-duplex-bench"}
{"slug":"humdial","name":"HumDial (ICASSP '26)","maintainer":"ICASSP 2026 Spoken Dialogue Challenge","tagline":"Dual-track human-evaluated spoken dialogue challenge — first public benchmark to explicitly score paralinguistic input alongside turn-taking dynamics.","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex","speech-lm","emotion"],"metrics":["turn-taking","paralinguistic grounding","human preference"],"models_scored":1,"updated":"2026-01","url":"https://aslp-lab.github.io/HumDial-Challenge/","paper_url":"https://arxiv.org/abs/2601.05564","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/humdial"}
{"slug":"flexi","name":"FLEXI","maintainer":"Ge et al.","tagline":"Full-duplex benchmark that includes the model interrupting the user in emergencies, across six interaction scenarios.","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex"],"metrics":["latency","response quality","emergency interruption"],"models_scored":0,"updated":"2025-09","url":"https://github.com/ChristineCHEN274/FLEXI","paper_url":"https://arxiv.org/abs/2509.22243","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/flexi"}
{"slug":"audiobench","name":"AudioBench","maintainer":"AudioLLMs (I2R Singapore)","tagline":"Broad audio-understanding benchmark covering speech, music, and ambient sound reasoning.","full_duplex":false,"tier":"component","setting":"lab","focus":["audio-understanding"],"metrics":["per-track accuracy","overall"],"models_scored":0,"updated":"2024-07","url":"https://github.com/AudioLLMs/AudioBench","paper_url":"https://arxiv.org/abs/2406.16020","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/audiobench"}
{"slug":"dynamic-superb","name":"Dynamic-SUPERB","maintainer":"NTU et al.","tagline":"Instruction-tuned successor to SUPERB, with 180 tasks in Phase-2 across speech understanding and light paralinguistic input.","full_duplex":false,"tier":"component","setting":"lab","focus":["speech-lm","instruction-following"],"metrics":["per-task accuracy","composite score"],"models_scored":0,"updated":"2024-11","url":"https://github.com/dynamic-superb/dynamic-superb","paper_url":"https://arxiv.org/abs/2411.05361","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/dynamic-superb"}
{"slug":"voiceassistant-eval","name":"VoiceAssistant-Eval","maintainer":"Wang et al.","tagline":"10,497 examples in 13 task categories that test voice assistants on listening, speaking and viewing.","full_duplex":false,"tier":"component","setting":"lab","focus":["speech-lm","instruction-following"],"metrics":["content quality","speech quality","consistency"],"models_scored":0,"updated":"2025-09","url":"https://mathllm.github.io/VoiceAssistantEval/","paper_url":"https://arxiv.org/abs/2509.22651","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/voiceassistant-eval"}
{"slug":"sd-eval","name":"SD-Eval","maintainer":"Academic (SD-Eval authors)","tagline":"Baseline probe on paralinguistic input — does a speech-LM use tone, emotion, and speaker traits at all?","full_duplex":false,"tier":"component","setting":"lab","focus":["speech-lm","emotion"],"metrics":["paralinguistic-grounded accuracy","delta over ASR baseline"],"models_scored":1,"updated":"2024-06","url":"https://arxiv.org/abs/2406.13340","paper_url":"https://arxiv.org/abs/2406.13340","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/sd-eval"}
{"slug":"prosaudit","name":"ProsAudit","maintainer":"Academic (ProsAudit authors)","tagline":"Prosodic boundary detection benchmark — can the model perceive phrase / sentence boundaries from audio alone?","full_duplex":false,"tier":"component","setting":"lab","focus":["emotion","audio-understanding"],"metrics":["boundary F1","delta over text-only"],"models_scored":0,"updated":"2023-02","url":"https://arxiv.org/abs/2302.12057","paper_url":"https://arxiv.org/abs/2302.12057","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/prosaudit"}
{"slug":"vocalbench-zh","name":"VocalBench-zh","maintainer":"SJTU (OmniAgent)","tagline":"Mandarin version of VocalBench — 10 subsets, ~10k instances, 14 models evaluated at the Nov 2025 launch.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","audio-understanding"],"metrics":["ACC","LLM-judge","UTMOS","WER","EER"],"models_scored":0,"updated":"2025-11","url":"https://github.com/SJTU-OmniAgent/VocalBench-zh","paper_url":"https://arxiv.org/abs/2511.08230","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/vocalbench-zh"}
{"slug":"cs3-bench","name":"CS3-Bench","maintainer":"Academic (CS3-Bench authors)","tagline":"Mandarin-English code-switching benchmark: across seven speech-to-speech models, knowledge-intensive question answering drops by up to 66% relative on code-switched input.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm"],"metrics":["monolingual-vs-codeswitch delta","per-pair accuracy"],"models_scored":0,"updated":"2025-10","url":"https://huggingface.co/datasets/VocalNet/CS3-Bench","paper_url":"https://arxiv.org/abs/2510.07881","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/cs3-bench"}
{"slug":"j-moshi","name":"J-Moshi (subjective)","maintainer":"Ohashi et al.","tagline":"The first public Japanese full-duplex dialogue model, built on Moshi and rated by human listeners for naturalness and meaningfulness.","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex","speech-lm"],"metrics":["naturalness","meaningfulness"],"models_scored":1,"updated":"2025-06","url":"https://nu-dialogue.github.io/j-moshi","paper_url":"https://arxiv.org/abs/2506.02979","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/j-moshi"}
{"slug":"voiceagentbench","name":"VoiceAgentBench","maintainer":"Academic (VoiceAgentBench authors)","tagline":"Spoken agentic benchmark: 6,000+ synthetic spoken queries covering single-tool calls, multi-tool workflows, multi-turn dialogue and safety, in English and six Indic languages.","full_duplex":false,"tier":"native","setting":"vertical","focus":["vertical-agent","speech-lm","instruction-following"],"metrics":["task success","tool-selection accuracy","multi-turn stability","safety"],"models_scored":0,"updated":"2026-02","url":"https://github.com/ola-krutrim/VoiceAgentBench","paper_url":"https://arxiv.org/abs/2510.07978","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/voiceagentbench"}
{"slug":"audio-multichallenge","name":"Audio MultiChallenge","maintainer":"Scale AI (Scale Labs)","tagline":"Multi-turn evaluation of spoken dialogue systems on natural human interaction — scores Inference Memory, Instruction Retention, Self Coherence, Voice Editing, Audio-Cue handling.","full_duplex":true,"tier":"native","setting":"lab","focus":["speech-lm","full-duplex"],"metrics":["Inference Memory","Instruction Retention","Self Coherence","Voice Editing","Audio-Cue"],"models_scored":4,"updated":"2026-10","url":"https://labs.scale.com/papers/audiomc","paper_url":"https://arxiv.org/abs/2512.14865","leaderboard_url":"https://labs.scale.com/leaderboard/audiomc","page_url":"https://www.fullduplex.ai/benchmarks/audio-multichallenge"}
{"slug":"msi-bench","name":"MSI-Bench","maintainer":"Boson AI","tagline":"Multi-speaker voice-agent benchmark: short multi-party audio scenes with participant context, expected tool calls and atomic rubrics, in English and Mandarin.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","vertical-agent"],"metrics":["all-rubric pass rate","multi-speaker memory","multi-speaker instruction following","multi-speaker reasoning"],"models_scored":0,"updated":"2026-09","url":"https://github.com/boson-ai/MSI-Bench","paper_url":"https://arxiv.org/abs/2609.24812","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/msi-bench"}
{"slug":"mmedfd","name":"MMedFD","maintainer":"Academic (MMedFD authors)","tagline":"Real-world healthcare benchmark for multi-turn full-duplex automatic speech recognition — adjacent to STS, crucial for duplex deployment.","full_duplex":true,"tier":"adjacent","setting":"vertical","focus":["asr","vertical-agent","full-duplex"],"metrics":["duplex WER","medical-term recall","overlap-region WER"],"models_scored":0,"updated":"2025-09","url":"https://github.com/Kinetics-JOJO/MMedFD","paper_url":"https://arxiv.org/abs/2509.19817","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/mmedfd"}
{"slug":"audio-jailbreaks-taxonomy","name":"LALM Jailbreak Taxonomy & Eval","maintainer":"Feng et al.","tagline":"Unified taxonomy of LALM jailbreak attacks (semantic, acoustic, signal, internal-representation) with controlled empirical evaluation under a single threat model and cost-aware protocol.","full_duplex":false,"tier":"native","setting":"lab","focus":["audio-safety","speech-lm"],"metrics":["attack success rate","cost-aware ASR","defense utility","cross-attack generalisation"],"models_scored":0,"updated":"2026-05","url":"https://arxiv.org/abs/2605.30031","paper_url":"https://arxiv.org/abs/2605.30031","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/audio-jailbreaks-taxonomy"}
{"slug":"speechjbb","name":"SpeechJBB","maintainer":"Ceccatelli, Jeon, Adelani et al.","tagline":"Audio jailbreak benchmark for LALMs under multilingual and code-switched speech, with phonological perturbations as an augmented evaluation regime.","full_duplex":false,"tier":"native","setting":"lab","focus":["audio-safety","speech-lm"],"metrics":["code-switched ASR","phonological-perturbation success","multilingual alignment gap"],"models_scored":0,"updated":"2026-08","url":"https://arxiv.org/abs/2606.06037","paper_url":"https://arxiv.org/abs/2606.06037","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/speechjbb"}
{"slug":"aia-acoustic-interference","name":"AIA — Acoustic Interference Attack","maintainer":"Wang, Huang, Liang, Wu (CUHK · HKUST)","tagline":"Universal-jailbreak benchmark for LALMs using Acoustic Latent Semantics — paralinguistic features intrinsic to the priors of audio generative models — that decouples attack payload from audio content.","full_duplex":false,"tier":"native","setting":"lab","focus":["audio-safety","speech-lm"],"metrics":["universal jailbreak success","alignment-evasion rate","transferability across LALMs"],"models_scored":0,"updated":"2026-05","url":"https://flaai.github.io/AIA_page/","paper_url":"https://arxiv.org/abs/2605.18168","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/aia-acoustic-interference"}
{"slug":"omni-duplexeval","name":"Omni-DuplexEval","maintainer":"OpenBMB (Tsinghua University / Shanghai Qi Zhi Institute / ModelBest)","tagline":"Streaming video benchmark that scores what a duplex omni-modal model says and when it says it, in live description and proactive reminder tasks.","full_duplex":true,"tier":"adjacent","setting":"lab","focus":["full-duplex","audio-understanding"],"metrics":["overall score (0 to 100)","Content Consistency","Temporal Sensitivity","Proactive Reminder score"],"models_scored":0,"updated":"2026-07","url":"https://github.com/OpenBMB/Omni-DuplexEval","paper_url":"https://arxiv.org/abs/2605.17360","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/omni-duplexeval"}
{"slug":"omniinteract","name":"OmniInteract","maintainer":"CUHK MMLab / SJTU / NTU / McMaster / CityUHK / JUFE","tagline":"Streaming audio-visual benchmark where spoken queries sit in the audio track and the model must decide whether, when and what to answer.","full_duplex":true,"tier":"native","setting":"live","focus":["full-duplex","audio-understanding"],"metrics":["IA-QTF1 (Interaction-Aware Quality-Timeliness F1)","Interruption Diagnostic Suite (NOR, PAQ, CSM)","Nested Chain Completion Score"],"models_scored":0,"updated":"2026-05","url":"https://github.com/Lucky-Lance/OmniInteract","paper_url":"https://arxiv.org/abs/2605.26485","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/omniinteract"}
{"slug":"provoice-bench","name":"ProVoice-Bench","maintainer":"Shanghai Jiao Tong University (Xu et al.)","tagline":"Tests whether a voice agent stays silent or steps in unprompted, across four proactive tasks built from synthesized conversations and device context.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","audio-understanding"],"metrics":["trigger accuracy","recall","false-positive rate","response accuracy (R_acc)"],"models_scored":0,"updated":"2026-06","url":"https://huggingface.co/datasets/overji/ProVoice-Bench","paper_url":"https://arxiv.org/abs/2604.15037","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/provoice-bench"}
{"slug":"eot-bench","name":"eot-bench","maintainer":"LiveKit","tagline":"End-of-turn detection scored at real pauses in human-to-agent turns across 14 languages, as a trade-off between false cutoffs and response delay.","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex"],"metrics":["false cutoffs at 300 ms and 600 ms latency budgets","latency at 5% and 10% false-cutoff budgets","latency and false-cutoff Pareto frontier"],"models_scored":0,"updated":"2026-11","url":"https://github.com/livekit/eot-bench","paper_url":null,"leaderboard_url":"https://livekit.com/benchmarks/eot-bench","page_url":"https://www.fullduplex.ai/benchmarks/eot-bench"}
{"slug":"audio2tool","name":"Audio2Tool","maintainer":"Rivian and Volkswagen Group Technologies","tagline":"Spoken tool-calling benchmark for smart car, smart home and wearable assistants, in eight tiers from direct commands to overlapping speakers.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","vertical-agent"],"metrics":["tool accuracy","exact match","parameter F1"],"models_scored":0,"updated":"2026-09","url":"https://audio2tool.github.io/","paper_url":"https://arxiv.org/abs/2604.22821","leaderboard_url":"https://huggingface.co/spaces/RVtech/Audio2ToolLeaderboard","page_url":"https://www.fullduplex.ai/benchmarks/audio2tool"}
{"slug":"c3-benchmark","name":"C3 Benchmark","maintainer":"Ma et al.","tagline":"English and Chinese test of whether spoken dialogue models handle ambiguity, omission, coreference and multi-turn context in speech.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm"],"metrics":["accuracy (LLM-judged)","per-phenomenon accuracy"],"models_scored":0,"updated":"2025-10","url":"https://step-out.github.io/C3-web/","paper_url":"https://arxiv.org/abs/2507.22968","leaderboard_url":"https://step-out.github.io/C3-web/","page_url":"https://www.fullduplex.ai/benchmarks/c3-benchmark"}
{"slug":"cekura-s2s-agent-bench","name":"Cekura Speech-to-Speech Agent Benchmark","maintainer":"Cekura","tagline":"Vendor-run leaderboard that swaps realtime speech models into one voice agent and scores 82 clinic and Medicare calls, three runs each.","full_duplex":false,"tier":"native","setting":"vertical","focus":["vertical-agent","speech-lm"],"metrics":["pass³ (all three runs pass)","success rate","data accuracy","median response time"],"models_scored":0,"updated":"2026-11","url":"https://benchmarks.cekura.ai/speech-to-speech","paper_url":null,"leaderboard_url":"https://benchmarks.cekura.ai/speech-to-speech","page_url":"https://www.fullduplex.ai/benchmarks/cekura-s2s-agent-bench"}
{"slug":"cekura-turn-bench","name":"Cekura Turn-Taking Benchmark","maintainer":"Cekura","tagline":"Scripted calls that test whether realtime speech models wait through pauses, reply at turn ends, stop on cut-ins and keep talking through backchannels.","full_duplex":true,"tier":"native","setting":"live","focus":["full-duplex"],"metrics":["pause handling","turn taking","barge-in yield","backchannel continue rate","end-of-turn latency"],"models_scored":0,"updated":"2026-11","url":"https://benchmarks.cekura.ai/speech-to-speech/turn-bench","paper_url":null,"leaderboard_url":"https://benchmarks.cekura.ai/speech-to-speech/turn-bench","page_url":"https://www.fullduplex.ai/benchmarks/cekura-turn-bench"}
{"slug":"codett","name":"CoDeTT","maintainer":"Bairong (BRVoice team)","tagline":"Turn-taking as a four-way decision with 14 labelled intents, so a model that picks the right action for the wrong reason is caught.","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex"],"metrics":["4-action accuracy (ACC)","Semantic Misalignment Rate (SMR)"],"models_scored":0,"updated":"2026-04","url":"https://huggingface.co/datasets/YingaoWang-casia/CoDeTT","paper_url":"https://arxiv.org/abs/2603.25434","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/codett"}
{"slug":"contextdialog","name":"ContextDialog","maintainer":"Seoul National University (Kim et al.)","tagline":"Spoken questions about something said earlier in a conversation, testing whether voice interaction models recall and use past turns.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm"],"metrics":["GPT Score (five-point)","word error rate"],"models_scored":0,"updated":"2025-05","url":"https://huggingface.co/datasets/ContextDialog/ContextDialog","paper_url":"https://arxiv.org/abs/2502.19759","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/contextdialog"}
{"slug":"coval-benchmarks","name":"Coval Benchmarks","maintainer":"Coval","tagline":"Continuously run vendor boards for speech-to-text, text-to-speech and speech-to-speech models, with an open-source runner for the first two and a dated methodology.","full_duplex":false,"tier":"component","setting":"live","focus":["asr","tts","speech-lm"],"metrics":["WER","TTFS","TTFA","voice-to-voice latency","instruction adherence"],"models_scored":0,"updated":"2026-11","url":"https://benchmarks.coval.ai/overview","paper_url":null,"leaderboard_url":"https://benchmarks.coval.ai/overview","page_url":"https://www.fullduplex.ai/benchmarks/coval-benchmarks"}
{"slug":"duplexworld","name":"DuplexWorld","maintainer":"Centific / University of Maryland","tagline":"Speech-to-speech voice agents on simulated telephone tasks in six worlds, scored on task completion, conversational dynamics and speech naturalness.","full_duplex":true,"tier":"native","setting":"vertical","focus":["vertical-agent","full-duplex"],"metrics":["Pass@1","goal-state check","turn-taking","selectivity","DNSMOS"],"models_scored":0,"updated":"2026-08","url":"https://huggingface.co/datasets/CentificAIResearch/Duplex-World","paper_url":"https://arxiv.org/abs/2608.10716","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/duplexworld"}
{"slug":"game-time","name":"Game-Time Benchmark","maintainer":"MIT / NTU / Academia Sinica","tagline":"Tests whether full-duplex spoken language models can follow spoken instructions about timing, tempo and simultaneous speech.","full_duplex":true,"tier":"native","setting":"lab","focus":["full-duplex","instruction-following"],"metrics":["instruction following (1 to 5)","response appropriateness (1 to 5)","turn-taking (1 to 5)"],"models_scored":0,"updated":"2026-06","url":"https://github.com/ga642381/Game-Time-Benchmark","paper_url":"https://arxiv.org/abs/2509.26388","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/game-time"}
{"slug":"ihbench","name":"IHBench","maintainer":"Boson AI","tagline":"Scores what a voice agent says after the user interrupts it in the middle of a structured workflow.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","vertical-agent"],"metrics":["task fulfillment win rate","recovery quality pass rate"],"models_scored":0,"updated":"2026-06","url":"https://huggingface.co/datasets/bosonai/IHBench","paper_url":"https://arxiv.org/abs/2606.19595","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/ihbench"}
{"slug":"kvoicebench","name":"KVoiceBench","maintainer":"KRAFTON","tagline":"Korean spoken question answering benchmark for speech language models, transferred from VoiceBench with human-agent translation and normalization rulebooks.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","instruction-following"],"metrics":["accuracy","judge score (open-ended)","prompt and instruction accuracy (KIFEval)","refusal rate (KAdvBench)"],"models_scored":0,"updated":"2026-08","url":"https://huggingface.co/datasets/KRAFTON/KVoiceBench","paper_url":"https://arxiv.org/abs/2605.27984","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/kvoicebench"}
{"slug":"mivas-bench","name":"MIVAS Bench","maintainer":"Bluejay","tagline":"Speech-to-speech models placed inside multi-agent call environments for healthcare, legal services and customer support, scored by deterministic verifiers.","full_duplex":false,"tier":"native","setting":"vertical","focus":["vertical-agent","speech-lm"],"metrics":["Pass^5","Pass@1","tool adherence","handoff adherence","final database state"],"models_scored":0,"updated":"2026-11","url":"https://github.com/bluejay-ai-dev/mivas-bench","paper_url":null,"leaderboard_url":"https://research.getbluejay.ai/benchmarks/mivas/leaderboard","page_url":"https://www.fullduplex.ai/benchmarks/mivas-bench"}
{"slug":"openaudiobench","name":"OpenAudioBench","maintainer":"Baichuan Inc.","tagline":"Spoken question answering set from Baichuan covering logical reasoning, general knowledge and open-ended questions, judged by GPT-4o.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm"],"metrics":["accuracy (Llama Questions, Web Questions, TriviaQA)","score (Reasoning QA, AlpacaEval)"],"models_scored":0,"updated":"2025-01","url":"https://huggingface.co/datasets/baichuan-inc/OpenAudioBench","paper_url":"https://arxiv.org/abs/2501.15368","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/openaudiobench"}
{"slug":"rolebreak","name":"RoleBreak","maintainer":"University of Hong Kong / Kami AI","tagline":"Tests whether speech-to-speech models keep a persona, stay safe and express emotion over role-play conversations of about 22 turns.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","emotion"],"metrics":["persona pass rate","interaction pass rate","safety pass rate","first-failure turn (P-FFT, S-FFT)","naturalness","emotion"],"models_scored":0,"updated":"2026-09","url":"https://huggingface.co/datasets/Greenbean/RoleBreak","paper_url":"https://arxiv.org/abs/2609.16614","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/rolebreak"}
{"slug":"s2s-arena","name":"S2S-Arena","maintainer":"FreedomIntelligence / CUHK-SZ / Shenzhen University of Advanced Technology","tagline":"Arena-style pairwise comparison of speech-to-speech models on spoken instructions that depend on paralinguistic cues, judged directly on audio.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","instruction-following","emotion"],"metrics":["Elo","win rate"],"models_scored":0,"updated":"2026-06","url":"https://github.com/FreedomIntelligence/S2S-Arena","paper_url":"https://arxiv.org/abs/2503.05085","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/s2s-arena"}
{"slug":"speechinstructbench","name":"SpeechInstructBench","maintainer":"CUHK / Alibaba Group","tagline":"Spoken instruction-following benchmark in English and Chinese with closed-ended, open-ended and adjustment tasks under accent, noise and disfluency variants.","full_duplex":false,"tier":"native","setting":"lab","focus":["instruction-following","speech-lm"],"metrics":["prompt-level accuracy","instruction-level accuracy","IAR","ECR"],"models_scored":0,"updated":"2026-01","url":"https://github.com/dingdongwang/SpeechInstructBench","paper_url":"https://arxiv.org/abs/2503.02769","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/speechinstructbench"}
{"slug":"steerbench","name":"SteerBench","maintainer":"Scale AI / University of Maryland","tagline":"Tests whether a speech-to-speech assistant changes its tone, persona, style or accent, and speed or length when the user asks for it out loud.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","instruction-following"],"metrics":["audio APR","sample APR","rubric pass rate"],"models_scored":0,"updated":"2026-11","url":"https://huggingface.co/datasets/utkarsh4430/SteerBench","paper_url":"https://arxiv.org/abs/2609.12623","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/steerbench"}
{"slug":"televal","name":"TELEVAL","maintainer":"China Telecom AI (Tele-AI)","tagline":"Chinese benchmark that checks whether spoken language models answer correctly and react appropriately to cues in the user's voice, without being told what to listen for.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","audio-understanding","emotion"],"metrics":["average score","Reliable Content Fulfillment","Interactional Appropriateness","speech-text consistency","DNSMOS"],"models_scored":0,"updated":"2026-08","url":"https://github.com/Tele-AI/TELEVAL","paper_url":"https://arxiv.org/abs/2507.18061","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/televal"}
{"slug":"tpi-bench","name":"TPI-Bench","maintainer":"Seoul National University / Yonsei University / University of Seoul","tagline":"Checks whether a voice assistant notices that a second person cut in on the user, and whether it then uses or ignores what that person said.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","audio-understanding"],"metrics":["Response Strategy Following (RSF)","Overall Helpfulness (OH)","BLEU","ROUGE-L"],"models_scored":0,"updated":"2026-04","url":"https://tpi-va.github.io/","paper_url":"https://arxiv.org/abs/2604.17358","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/tpi-bench"}
{"slug":"vera","name":"VERA","maintainer":"Duke University / Adobe","tagline":"Measures how much reasoning accuracy a model loses when the same questions are asked and answered by voice under real-time limits.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm"],"metrics":["accuracy (macro-averaged over tracks)","time to first response (TTFR)","WER"],"models_scored":0,"updated":"2025-09","url":"https://github.com/linyueqian/VERA","paper_url":"https://arxiv.org/abs/2509.26542","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/vera"}
{"slug":"voicebench-ja","name":"VoiceBench-ja","maintainer":"SB Intuitions","tagline":"Japanese test set for the gap between spoken and typed input, with the same prompts as audio and as text in four subsets.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","instruction-following"],"metrics":["LLM-judge score (1 to 5)","format following","exact match"],"models_scored":0,"updated":"2026-03","url":"https://huggingface.co/datasets/sbintuitions/voicebench-ja","paper_url":"https://arxiv.org/abs/2603.12565","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/voicebench-ja"}
{"slug":"voxeval","name":"VoxEval","maintainer":"CUHK / Lightspeed Studios / NUS","tagline":"MMLU in spoken form: multiple-choice knowledge questions asked and answered in speech, repeated across speakers, speaking styles and audio quality.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm"],"metrics":["accuracy (ASR + string matching)"],"models_scored":0,"updated":"2025-07","url":"https://github.com/dreamtheater123/VoxEval","paper_url":"https://arxiv.org/abs/2501.04962","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/voxeval"}
{"slug":"wavbench","name":"WavBench","maintainer":"Xiamen University / Zhejiang University / CUHK-Shenzhen","tagline":"Scores spoken dialogue models on hard reasoning, on whether answers sound like natural talk, and on hearing and producing paralinguistic cues.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","emotion","audio-understanding"],"metrics":["colloquial score (Pro, Basic)","explicit understanding accuracy","explicit generation accuracy","implicit dialogue score"],"models_scored":0,"updated":"2026-02","url":"https://naruto-2024.github.io/wavbench.github.io/","paper_url":"https://arxiv.org/abs/2602.12135","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/wavbench"}
{"slug":"wearvox","name":"WearVox","maintainer":"Meta","tagline":"Voice assistant benchmark recorded on AI glasses, with 3,842 multi-channel egocentric clips that include real indoor and outdoor noise and bystander speech.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm","audio-understanding"],"metrics":["turn-based micro-average accuracy","tool-call AST accuracy","side-talk rejection accuracy","speech translation judge score"],"models_scored":0,"updated":"2026-04","url":"https://github.com/facebookresearch/WearVox","paper_url":"https://arxiv.org/abs/2601.02391","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/wearvox"}
{"slug":"wildspeech-bench","name":"WildSpeech-Bench","maintainer":"Tencent (WeChat AI)","tagline":"Speech-to-speech test built from real user chat queries, with varied speakers, added noise, and a human-recorded set of pauses, stress, tone, stuttering and near-homophones.","full_duplex":false,"tier":"native","setting":"lab","focus":["speech-lm"],"metrics":["average score (1 to 10)","per-category score","UTMOS"],"models_scored":0,"updated":"2025-09","url":"https://github.com/Tencent/WildSpeech-Bench","paper_url":"https://arxiv.org/abs/2506.21875","leaderboard_url":null,"page_url":"https://www.fullduplex.ai/benchmarks/wildspeech-bench"}