slug stringlengths 4 25 | name stringlengths 4 42 | maintainer stringlengths 3 74 | tagline stringlengths 69 200 | full_duplex bool 2
classes | tier stringclasses 4
values | setting stringclasses 4
values | focus listlengths 1 4 | metrics listlengths 1 6 | models_scored int64 0 24 | updated stringdate 2020-07-01 00:00:00 2026-11-01 00:00:00 | url stringlengths 23 80 | paper_url stringlengths 32 32 ⌀ | leaderboard_url stringlengths 24 73 ⌀ | page_url stringlengths 41 62 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
turnbench | TurnBench | Sesame AI Labs / Mundo AI / CMU / oto | Scores end-of-turn and interruption detection on 30 hours of hand-labelled dyadic conversation, with conversation type as a controlled variable. | true | native | lab | [
"full-duplex"
] | [
"end-of-turn recall",
"interruption recall",
"false-positive rate",
"median detection latency"
] | 3 | 2026-11 | https://turnbench.sesame.com/ | https://arxiv.org/abs/2608.25218 | https://turnbench.sesame.com/ | https://www.fullduplex.ai/benchmarks/turnbench |
fdb-v1 | Full-Duplex-Bench v1 | NTU / UC Berkeley / UW / MIT | Original Full-Duplex-Bench — automatic scoring of the four canonical turn-taking events: when-to-speak, back-channel, interruption success, pause handling. | true | native | lab | [
"full-duplex"
] | [
"turn-taking latency",
"back-channel timing",
"interruption success",
"pause handling"
] | 14 | 2025-03 | https://github.com/DanielLin94144/Full-Duplex-Bench | https://arxiv.org/abs/2503.04721 | null | https://www.fullduplex.ai/benchmarks/fdb-v1 |
fdb-v15 | Full-Duplex-Bench v1.5 | NTU / UC Berkeley / UW / MIT | Mid-2025 refresh of FDB v1 — adds the simultaneous-speech / overlap axis (user interruption, listener back-channel, side conversation, ambient speech). | true | native | lab | [
"full-duplex"
] | [
"overlap duration",
"mutual-silence rate",
"interruption recovery"
] | 6 | 2025-08 | https://github.com/DanielLin94144/Full-Duplex-Bench | https://arxiv.org/abs/2507.23159 | null | https://www.fullduplex.ai/benchmarks/fdb-v15 |
fdb-v2 | Full-Duplex-Bench v2 | NTU / UC Berkeley / UW / MIT | Late-2025 WebRTC-based live harness — first FDB release that runs dynamic evaluation against streaming endpoints, plus multi-turn stimuli. | true | native | live | [
"full-duplex",
"speech-lm"
] | [
"streaming latency",
"multi-turn coherence",
"overlap handling"
] | 2 | 2026-02 | https://github.com/DanielLin94144/Full-Duplex-Bench | https://arxiv.org/abs/2510.07838 | null | https://www.fullduplex.ai/benchmarks/fdb-v2 |
fdb-v3 | Full-Duplex-Bench v3 | NTU (NVIDIA advisory) | Multi-step tool calling for voice agents, driven by real human speech with natural disfluencies. | true | native | live | [
"full-duplex",
"vertical-agent"
] | [
"Pass@1",
"tool-call F1",
"response latency",
"turn-take rate"
] | 6 | 2026-05 | https://github.com/DanielLin94144/Full-Duplex-Bench | https://arxiv.org/abs/2604.04847 | null | https://www.fullduplex.ai/benchmarks/fdb-v3 |
videofdb | VideoFDB | NVIDIA / David AI | 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? | true | native | live | [
"full-duplex"
] | [
"perception rubric (0 to 5)",
"generation rubric (0 to 5)",
"TOR-Alignment",
"median response latency"
] | 4 | 2026-10 | https://research.nvidia.com/labs/amri/projects/video-fdb/ | https://arxiv.org/abs/2605.30256 | https://research.nvidia.com/labs/amri/projects/video-fdb/leaderboard.html | https://www.fullduplex.ai/benchmarks/videofdb |
duplexact | DuplexAct | Dong et al. | Scores whether full-duplex speech models take the right interaction action when a conversation calls for it, and whether they follow through. | true | native | lab | [
"full-duplex"
] | [
"immediate execution success",
"follow-through success"
] | 8 | 2026-10 | https://github.com/JiahengDong/DuplexAct | https://arxiv.org/abs/2610.17738 | null | https://www.fullduplex.ai/benchmarks/duplexact |
multipartyturnbench | MultiPartyTurnBench | Mathur and Manocha | 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? | true | native | lab | [
"full-duplex"
] | [
"end-of-turn recall at fixed FPR",
"interruption recall at fixed FPR",
"signed latency",
"selective interruption balanced accuracy"
] | 2 | 2026-10 | https://multipartyturnbench.github.io/ | https://arxiv.org/abs/2610.20172 | null | https://www.fullduplex.ai/benchmarks/multipartyturnbench |
multitalkbench | MultiTalkBench | Wang et al. (MultiTalk) | Long, multi-party, English and Chinese full-duplex dialogue, with conversations that average 32.6 minutes. | true | native | lab | [
"full-duplex"
] | [
"long-range entity tracking",
"topic coherence",
"addressee selection",
"turn-taking"
] | 0 | 2026-09 | https://huggingface.co/datasets/MultiTalk/MultiTalkBench | https://arxiv.org/abs/2609.36903 | null | https://www.fullduplex.ai/benchmarks/multitalkbench |
m3-duplexbench | M3-DuplexBench | NTT (Fukuda et al.) | English and Japanese full-duplex benchmark that scores interaction timing and response content together, in casual chat and multi-turn QA. | true | native | lab | [
"full-duplex",
"speech-lm"
] | [
"turn-taking rate and latency",
"pause handling",
"backchannel and barge-in response",
"response content"
] | 2 | 2026-10 | https://github.com/nttcslab-sp/m3-duplexbench | https://arxiv.org/abs/2607.29125 | null | https://www.fullduplex.ai/benchmarks/m3-duplexbench |
duplex-mpe | Duplex-MPE | Ma et al. | Multi-party full-duplex test: with three or four people talking, should the assistant answer, stay silent or stop? | true | native | lab | [
"full-duplex"
] | [
"response initiation",
"answer accuracy",
"silence preservation",
"stopping when resolved"
] | 0 | 2026-11 | https://step-out.github.io/Duplex-MPE-Page/ | https://arxiv.org/abs/2609.31948 | null | https://www.fullduplex.ai/benchmarks/duplex-mpe |
duplexledger | DuplexLedger | Li et al. | Does a full-duplex agent remember the right things? Write, update, ignore and retain decisions during live interaction. | true | native | lab | [
"full-duplex",
"speech-lm"
] | [
"latest-value accuracy",
"final-state fact coverage",
"stale assertions",
"distractor adoption"
] | 0 | 2026-10 | https://github.com/lyx-researcher/DuplexLedger | https://arxiv.org/abs/2610.27581 | null | https://www.fullduplex.ai/benchmarks/duplexledger |
interactionbench | InteractionBench | Avera Labs (He et al.) | Timing, spoken-task accuracy, whisper and volume, and groundedness over 20+ turns, measured on the same recorded user audio for every system. | true | native | live | [
"full-duplex",
"speech-lm"
] | [
"conversational timing",
"spoken-task accuracy",
"whisper and volume control",
"groundedness"
] | 0 | 2026-11 | https://avera-labs.github.io/InteractionBench/ | https://arxiv.org/abs/2611.02430 | null | https://www.fullduplex.ai/benchmarks/interactionbench |
echo-turn-taking | ECHO | Zhao et al. | Paired Chinese turn-taking test: the same overlapping words must sometimes take the floor and sometimes not, depending on context. | true | native | lab | [
"full-duplex"
] | [
"pair accuracy",
"yield accuracy",
"keep accuracy",
"off-talk false yields"
] | 0 | 2026-09 | https://huggingface.co/datasets/shuofeng123/ECHO | https://arxiv.org/abs/2609.17360 | null | https://www.fullduplex.ai/benchmarks/echo-turn-taking |
mp-bench | MP-Bench | Shih et al. | The voice agent is one participant in a spoken group conversation: does it follow who said what, and does it take its turn? | true | native | lab | [
"full-duplex",
"speech-lm"
] | [
"turn-taking accuracy",
"response appropriateness",
"comprehension accuracy",
"latency"
] | 0 | 2026-10 | https://github.com/atosystem/MP-Bench | https://arxiv.org/abs/2609.13076 | null | https://www.fullduplex.ai/benchmarks/mp-bench |
dsb-ifeval | DuplexSpeechBench-IFEval | Mathur and Manocha | Does a full-duplex agent infer from its persona when to speak, backchannel, interrupt or yield, and not only from explicit rules? | true | native | lab | [
"full-duplex",
"instruction-following"
] | [
"Instruction Adherence Score",
"Persona Adherence Score"
] | 0 | 2026-09 | https://dsb-ifeval.github.io/ | https://arxiv.org/abs/2609.03423 | null | https://www.fullduplex.ai/benchmarks/dsb-ifeval |
talking-turns | Talking Turns | Apple / CMU | A supervised turn-taking judge trained on human-human conversations to score audio foundation models. | true | native | lab | [
"full-duplex",
"speech-lm"
] | [
"turn-event F1",
"backchannel rate",
"interruption rate"
] | 3 | 2025-04 | https://machinelearning.apple.com/research/talking-turns | https://arxiv.org/abs/2503.01174 | null | https://www.fullduplex.ai/benchmarks/talking-turns |
voicebench | VoiceBench | NUS | Evaluates LLM-based voice assistants across knowledge, reasoning, safety, and instruction-following — judged by a GPT-4-class model. | false | native | lab | [
"speech-lm",
"instruction-following"
] | [
"overall",
"QA",
"reasoning (BBH)",
"IFEval",
"safety"
] | 24 | 2025-06 | https://github.com/MatthewCYM/VoiceBench | https://arxiv.org/abs/2410.17196 | https://matthewcym.github.io/VoiceBench/ | https://www.fullduplex.ai/benchmarks/voicebench |
vocalbench | VocalBench | SJTU (OmniAgent) | A four-dimensional stress test for speech-interaction models — semantic, acoustic, conversational, and robust. | false | native | lab | [
"speech-lm",
"audio-understanding"
] | [
"ACC",
"LLM-judge",
"UTMOS",
"WER",
"EER"
] | 1 | 2026-01 | https://github.com/SJTU-OmniAgent/VocalBench | https://arxiv.org/abs/2505.15727 | null | https://www.fullduplex.ai/benchmarks/vocalbench |
uro-bench | URO-Bench | Ruiqi Yan et al. | The first end-to-end S2S benchmark with multilingual, multi-turn, and paralinguistic coverage. | false | native | lab | [
"speech-lm",
"instruction-following",
"emotion"
] | [
"U/R/O composite",
"paralinguistic score"
] | 5 | 2025-08 | https://github.com/Ruiqi-Yan/URO-Bench | https://arxiv.org/abs/2502.17810 | null | https://www.fullduplex.ai/benchmarks/uro-bench |
air-bench | AIR-Bench | Alibaba DAMO (OFA-Sys) | A two-tier stress test for audio-LLMs covering speech, natural sound, and music — MCQ plus chat. | false | native | lab | [
"audio-understanding",
"instruction-following"
] | [
"MCQ acc",
"GPT-4 judge score"
] | 4 | 2024-12 | https://github.com/OFA-Sys/AIR-Bench | https://arxiv.org/abs/2402.07729 | null | https://www.fullduplex.ai/benchmarks/air-bench |
big-bench-audio | Big Bench Audio | Artificial Analysis | 1 000 audio-question reasoning benchmark — the dataset behind the Artificial Analysis S2S speech-reasoning leaderboard. | false | native | lab | [
"speech-lm",
"instruction-following"
] | [
"accuracy",
"TTFA (s)"
] | 12 | 2026-05 | https://huggingface.co/datasets/ArtificialAnalysis/big_bench_audio | null | https://artificialanalysis.ai/speech-to-speech | https://www.fullduplex.ai/benchmarks/big-bench-audio |
aa-speech-to-speech-index | Artificial Analysis Speech to Speech Index | Artificial Analysis | Composite score for native speech-to-speech models: speech reasoning, agentic task completion, human preference and task success, weighted equally. | false | native | lab | [
"speech-lm",
"vertical-agent"
] | [
"index score",
"Big Bench Audio accuracy",
"τ-Voice pass rate",
"Arena Score",
"task success rate"
] | 0 | 2026-08 | https://artificialanalysis.ai/methodology/speech-to-speech-benchmarking | null | https://artificialanalysis.ai/speech-to-speech | https://www.fullduplex.ai/benchmarks/aa-speech-to-speech-index |
mmar | MMAR | SJTU / MBZUAI · NeurIPS 2025 | 1 000 human-curated deep-reasoning audio QA items spanning speech, sound, music, and their mix — four hierarchical reasoning layers. | false | component | lab | [
"audio-understanding",
"speech-lm"
] | [
"overall accuracy",
"per-layer accuracy",
"rubric score"
] | 5 | 2026-02 | https://github.com/ddlBoJack/MMAR | https://arxiv.org/abs/2505.13032 | null | https://www.fullduplex.ai/benchmarks/mmar |
ifeval-audio | IFEval-Audio | AudioLLMs (I2R Singapore) | Format-constrained spoken instructions, verified programmatically — an audio port of text IFEval. | false | native | lab | [
"instruction-following",
"speech-lm"
] | [
"IFR",
"SCR",
"OSR"
] | 0 | 2025-05 | https://github.com/AudioLLMs/AudioBench/tree/main/IFEval-Audio | https://arxiv.org/abs/2505.16774 | null | https://www.fullduplex.ai/benchmarks/ifeval-audio |
scale-voice-showdown | Scale Voice Showdown | Scale AI | 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. | false | native | arena | [
"speech-lm",
"tts"
] | [
"preference win-rate",
"per-language score"
] | 1 | 2026-03 | https://labs.scale.com/ | null | null | https://www.fullduplex.ai/benchmarks/scale-voice-showdown |
aa-speech-arena | Artificial Analysis Speech Arena | Artificial Analysis | 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. | false | adjacent | arena | [
"tts"
] | [
"Elo"
] | 0 | 2026-10 | https://artificialanalysis.ai/text-to-speech/arena | null | https://artificialanalysis.ai/text-to-speech/leaderboard | https://www.fullduplex.ai/benchmarks/aa-speech-arena |
aa-speech-agent-arena | Artificial Analysis Speech Agent Arena | Artificial Analysis | Human participants complete the same task with two hidden speech-to-speech systems over live calls, then vote for the conversation they preferred. | false | native | arena | [
"speech-lm",
"vertical-agent"
] | [
"preference Elo",
"agentic Elo",
"non-agentic Elo",
"task success rate"
] | 0 | 2026-10 | https://artificialanalysis.ai/speech-to-speech/arena/overview | null | https://artificialanalysis.ai/speech-to-speech/arena | https://www.fullduplex.ai/benchmarks/aa-speech-agent-arena |
rw-voice-eq | Real World Voice EQ Bench | Hume AI | 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. | false | component | arena | [
"tts",
"speech-lm",
"asr",
"emotion"
] | [
"rater score (1 to 5) per factor",
"WER (ASR board)"
] | 0 | 2026-07 | https://huggingface.co/spaces/HumeAI/rw-voice-eq | https://arxiv.org/abs/2607.14846 | https://huggingface.co/spaces/HumeAI/rw-voice-eq | https://www.fullduplex.ai/benchmarks/rw-voice-eq |
mtalk-bench | MTalk-Bench | FreedomIntelligence / CUHK-SZ | First multi-turn S2S benchmark with both arena-style (pairwise) and rubric-based (absolute) protocols for dialogue evaluation. | false | native | arena | [
"speech-lm",
"instruction-following",
"emotion"
] | [
"arena win-rate",
"rubric score",
"per-dimension score"
] | 0 | 2025-09 | https://github.com/FreedomIntelligence/MTalk-Bench | https://arxiv.org/abs/2508.18240 | null | https://www.fullduplex.ai/benchmarks/mtalk-bench |
tts-arena | TTS Arena | Hugging Face | The original blind A/B voting space for TTS, feeding an Elo ladder on Hugging Face. | false | adjacent | arena | [
"tts"
] | [
"Elo"
] | 0 | 2026-03 | https://huggingface.co/spaces/TTS-AGI/TTS-Arena-V2 | null | https://huggingface.co/spaces/TTS-AGI/TTS-Arena-V2 | https://www.fullduplex.ai/benchmarks/tts-arena |
tau3-bench | τ-bench 2.0 (τ-Voice) | Sierra Research | The first agent benchmark with a live, full-duplex voice track, now in six languages with version 2.0. | true | native | vertical | [
"vertical-agent",
"full-duplex",
"speech-lm"
] | [
"pass^k",
"tool-use success",
"voice task success",
"per-language task completion"
] | 5 | 2026-10 | https://sierra.ai/blog/bench-advancing-agent-benchmarking-to-knowledge-and-voice | https://arxiv.org/abs/2603.13686 | https://www.taubench.com | https://www.fullduplex.ai/benchmarks/tau3-bench |
eva-bench | EVA-Bench | ServiceNow | End-to-end voice-agent evaluation with bot-to-bot calls, separate accuracy and experience scores, and accent and noise perturbations. | false | native | vertical | [
"vertical-agent",
"speech-lm"
] | [
"EVA-A (accuracy)",
"EVA-X (experience)",
"pass@k and pass^k",
"accent and noise robustness"
] | 1 | 2026-10 | https://servicenow.github.io/eva/ | https://arxiv.org/abs/2605.13841 | https://servicenow.github.io/eva/ | https://www.fullduplex.ai/benchmarks/eva-bench |
vamos-bench | VAmoS Bench | Veris AI | Simulated phone calls graded on what the voice agent said and on what it actually changed in the backend. | false | native | vertical | [
"vertical-agent"
] | [
"task completion",
"latency",
"barge-ins per call",
"cost per call"
] | 0 | 2026-10 | https://veris.ai/leaderboard/vamos | https://arxiv.org/abs/2609.38512 | https://veris.ai/leaderboard/vamos-pro | https://www.fullduplex.ai/benchmarks/vamos-bench |
apex-voice | APEX-Voice | Mathur and Manocha | Can a full-duplex voice agent finish professional work, such as form completion or negotiation, while the user interrupts and corrects? | true | native | vertical | [
"vertical-agent",
"full-duplex"
] | [
"Pass@1",
"Reliable@3",
"artifact field accuracy"
] | 0 | 2026-09 | https://apex-voice.github.io/ | https://arxiv.org/abs/2609.34973 | null | https://www.fullduplex.ai/benchmarks/apex-voice |
voiceagenteval | VoiceAgentEval | Meituan / Xbench / Agora | A dual-dimensional benchmark for expert-level outbound-calling agents — interaction fluency vs task flow compliance. | false | native | vertical | [
"vertical-agent",
"speech-lm"
] | [
"GIC",
"TFC",
"task success"
] | 0 | 2025-10 | https://www.shengwang.cn/news/blogdetail/VoiceAgentEval/ | https://arxiv.org/abs/2510.21244 | null | https://www.fullduplex.ai/benchmarks/voiceagenteval |
healthbench | HealthBench | OpenAI | 5,000 multi-turn health conversations × 48k physician-written rubric criteria — text-only, listed as a transferable foundation rubric rather than a voice benchmark. | false | adjacent | vertical | [
"vertical-agent",
"instruction-following"
] | [
"rubric score",
"per-theme breakdown"
] | 0 | 2025-05 | https://huggingface.co/datasets/openai/healthbench | https://arxiv.org/abs/2505.08775 | null | https://www.fullduplex.ai/benchmarks/healthbench |
seamless-expressive | SeamlessExpressive (mExpresso) | Meta AI | Expressive speech-to-speech translation judged on translation, prosody, and vocal-style preservation. | false | component | lab | [
"s2st"
] | [
"ASR-BLEU",
"AutoPCP",
"Vocal Style Sim"
] | 0 | 2023-12 | https://github.com/facebookresearch/seamless_communication | https://arxiv.org/abs/2312.05187 | null | https://www.fullduplex.ai/benchmarks/seamless-expressive |
covost2 | CoVoST-2 BLEU | Meta AI | 21 language pairs of speech-to-text translation, long the S2T standard. | false | component | lab | [
"s2st",
"asr"
] | [
"BLEU",
"chrF"
] | 5 | 2020-07 | https://github.com/facebookresearch/covost | https://arxiv.org/abs/2007.10310 | null | https://www.fullduplex.ai/benchmarks/covost2 |
seed-tts-eval | Seed-TTS Eval | ByteDance | The de-facto evaluation set for zero-shot voice cloning — WER, speaker similarity, and CMOS in one recipe. | false | component | lab | [
"tts"
] | [
"WER",
"SECS",
"CMOS"
] | 3 | 2024-11 | https://bytedancespeech.github.io/seedtts_tech_report/ | https://arxiv.org/abs/2406.02430 | null | https://www.fullduplex.ai/benchmarks/seed-tts-eval |
nv-bench | NV-Bench | Ni et al. | Standardised test for generating laughs, sighs, fillers, and back-channels in TTS. | false | component | lab | [
"tts",
"emotion"
] | [
"PCER",
"FAD",
"SIM",
"DNSMOS"
] | 0 | 2026-03 | https://arxiv.org/abs/2603.15352 | https://arxiv.org/abs/2603.15352 | null | https://www.fullduplex.ai/benchmarks/nv-bench |
superb | SUPERB | NTU et al. (s3prl) | The classic: general-purpose speech representations scored on ten frozen-encoder downstream tasks. | false | peripheral | lab | [
"speech-lm",
"asr",
"audio-understanding"
] | [
"composite score",
"per-task scores"
] | 0 | 2024-06 | https://github.com/s3prl/s3prl | https://arxiv.org/abs/2105.01051 | null | https://www.fullduplex.ai/benchmarks/superb |
librispeech | LibriSpeech WER | OpenSLR | The coordinate system of English ASR — test-clean and test-other WER. | false | peripheral | lab | [
"asr"
] | [
"WER (test-clean)",
"WER (test-other)"
] | 13 | 2024-01 | https://www.openslr.org/12 | null | null | https://www.fullduplex.ai/benchmarks/librispeech |
meld | MELD / IEMOCAP | Academic | The classic emotion-recognition pair, evaluated with full dialogue context. | false | peripheral | lab | [
"emotion",
"audio-understanding"
] | [
"Weighted F1",
"Accuracy"
] | 0 | 2023-06 | https://affective-meld.github.io | null | null | https://www.fullduplex.ai/benchmarks/meld |
sid-bench | SID-Bench | Xia et al. | Semantic-aware interruption detection, built from real human dialogues and scored with a time penalty for both false alarms and late responses. | true | native | lab | [
"full-duplex"
] | [
"Average Penalty Time (APT)",
"false alarms",
"response delay"
] | 0 | 2026-03 | https://github.com/xkx-hub/SID-bench | https://arxiv.org/abs/2603.24144 | null | https://www.fullduplex.ai/benchmarks/sid-bench |
fd-bench | FD-Bench | Peng et al. | A pipeline of LLM, TTS and ASR that simulates users and tests how full-duplex systems handle interruptions, delays and noise. | true | native | lab | [
"full-duplex"
] | [
"interruption response",
"delay",
"robustness under noise"
] | 0 | 2025-07 | https://github.com/pengyizhou/FD-Bench | https://arxiv.org/abs/2507.19040 | null | https://www.fullduplex.ai/benchmarks/fd-bench |
mtr-duplex-bench | MTR-DuplexBench | Zhang et al. | Multi-round benchmark that splits continuous full-duplex dialogue into turns and scores each round. | true | native | lab | [
"full-duplex",
"speech-lm"
] | [
"conversational features",
"dialogue quality",
"instruction following",
"safety"
] | 0 | 2026-04 | https://github.com/ZhangHe0918/MTR-DuplexBench | https://arxiv.org/abs/2511.10262 | null | https://www.fullduplex.ai/benchmarks/mtr-duplex-bench |
humdial | HumDial (ICASSP '26) | ICASSP 2026 Spoken Dialogue Challenge | Dual-track human-evaluated spoken dialogue challenge — first public benchmark to explicitly score paralinguistic input alongside turn-taking dynamics. | true | native | lab | [
"full-duplex",
"speech-lm",
"emotion"
] | [
"turn-taking",
"paralinguistic grounding",
"human preference"
] | 1 | 2026-01 | https://aslp-lab.github.io/HumDial-Challenge/ | https://arxiv.org/abs/2601.05564 | null | https://www.fullduplex.ai/benchmarks/humdial |
flexi | FLEXI | Ge et al. | Full-duplex benchmark that includes the model interrupting the user in emergencies, across six interaction scenarios. | true | native | lab | [
"full-duplex"
] | [
"latency",
"response quality",
"emergency interruption"
] | 0 | 2025-09 | https://github.com/ChristineCHEN274/FLEXI | https://arxiv.org/abs/2509.22243 | null | https://www.fullduplex.ai/benchmarks/flexi |
audiobench | AudioBench | AudioLLMs (I2R Singapore) | Broad audio-understanding benchmark covering speech, music, and ambient sound reasoning. | false | component | lab | [
"audio-understanding"
] | [
"per-track accuracy",
"overall"
] | 0 | 2024-07 | https://github.com/AudioLLMs/AudioBench | https://arxiv.org/abs/2406.16020 | null | https://www.fullduplex.ai/benchmarks/audiobench |
dynamic-superb | Dynamic-SUPERB | NTU et al. | Instruction-tuned successor to SUPERB, with 180 tasks in Phase-2 across speech understanding and light paralinguistic input. | false | component | lab | [
"speech-lm",
"instruction-following"
] | [
"per-task accuracy",
"composite score"
] | 0 | 2024-11 | https://github.com/dynamic-superb/dynamic-superb | https://arxiv.org/abs/2411.05361 | null | https://www.fullduplex.ai/benchmarks/dynamic-superb |
voiceassistant-eval | VoiceAssistant-Eval | Wang et al. | 10,497 examples in 13 task categories that test voice assistants on listening, speaking and viewing. | false | component | lab | [
"speech-lm",
"instruction-following"
] | [
"content quality",
"speech quality",
"consistency"
] | 0 | 2025-09 | https://mathllm.github.io/VoiceAssistantEval/ | https://arxiv.org/abs/2509.22651 | null | https://www.fullduplex.ai/benchmarks/voiceassistant-eval |
sd-eval | SD-Eval | Academic (SD-Eval authors) | Baseline probe on paralinguistic input — does a speech-LM use tone, emotion, and speaker traits at all? | false | component | lab | [
"speech-lm",
"emotion"
] | [
"paralinguistic-grounded accuracy",
"delta over ASR baseline"
] | 1 | 2024-06 | https://arxiv.org/abs/2406.13340 | https://arxiv.org/abs/2406.13340 | null | https://www.fullduplex.ai/benchmarks/sd-eval |
prosaudit | ProsAudit | Academic (ProsAudit authors) | Prosodic boundary detection benchmark — can the model perceive phrase / sentence boundaries from audio alone? | false | component | lab | [
"emotion",
"audio-understanding"
] | [
"boundary F1",
"delta over text-only"
] | 0 | 2023-02 | https://arxiv.org/abs/2302.12057 | https://arxiv.org/abs/2302.12057 | null | https://www.fullduplex.ai/benchmarks/prosaudit |
vocalbench-zh | VocalBench-zh | SJTU (OmniAgent) | Mandarin version of VocalBench — 10 subsets, ~10k instances, 14 models evaluated at the Nov 2025 launch. | false | native | lab | [
"speech-lm",
"audio-understanding"
] | [
"ACC",
"LLM-judge",
"UTMOS",
"WER",
"EER"
] | 0 | 2025-11 | https://github.com/SJTU-OmniAgent/VocalBench-zh | https://arxiv.org/abs/2511.08230 | null | https://www.fullduplex.ai/benchmarks/vocalbench-zh |
cs3-bench | CS3-Bench | Academic (CS3-Bench authors) | 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. | false | native | lab | [
"speech-lm"
] | [
"monolingual-vs-codeswitch delta",
"per-pair accuracy"
] | 0 | 2025-10 | https://huggingface.co/datasets/VocalNet/CS3-Bench | https://arxiv.org/abs/2510.07881 | null | https://www.fullduplex.ai/benchmarks/cs3-bench |
j-moshi | J-Moshi (subjective) | Ohashi et al. | The first public Japanese full-duplex dialogue model, built on Moshi and rated by human listeners for naturalness and meaningfulness. | true | native | lab | [
"full-duplex",
"speech-lm"
] | [
"naturalness",
"meaningfulness"
] | 1 | 2025-06 | https://nu-dialogue.github.io/j-moshi | https://arxiv.org/abs/2506.02979 | null | https://www.fullduplex.ai/benchmarks/j-moshi |
voiceagentbench | VoiceAgentBench | Academic (VoiceAgentBench authors) | 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. | false | native | vertical | [
"vertical-agent",
"speech-lm",
"instruction-following"
] | [
"task success",
"tool-selection accuracy",
"multi-turn stability",
"safety"
] | 0 | 2026-02 | https://github.com/ola-krutrim/VoiceAgentBench | https://arxiv.org/abs/2510.07978 | null | https://www.fullduplex.ai/benchmarks/voiceagentbench |
audio-multichallenge | Audio MultiChallenge | Scale AI (Scale Labs) | Multi-turn evaluation of spoken dialogue systems on natural human interaction — scores Inference Memory, Instruction Retention, Self Coherence, Voice Editing, Audio-Cue handling. | true | native | lab | [
"speech-lm",
"full-duplex"
] | [
"Inference Memory",
"Instruction Retention",
"Self Coherence",
"Voice Editing",
"Audio-Cue"
] | 4 | 2026-10 | https://labs.scale.com/papers/audiomc | https://arxiv.org/abs/2512.14865 | https://labs.scale.com/leaderboard/audiomc | https://www.fullduplex.ai/benchmarks/audio-multichallenge |
msi-bench | MSI-Bench | Boson AI | Multi-speaker voice-agent benchmark: short multi-party audio scenes with participant context, expected tool calls and atomic rubrics, in English and Mandarin. | false | native | lab | [
"speech-lm",
"vertical-agent"
] | [
"all-rubric pass rate",
"multi-speaker memory",
"multi-speaker instruction following",
"multi-speaker reasoning"
] | 0 | 2026-09 | https://github.com/boson-ai/MSI-Bench | https://arxiv.org/abs/2609.24812 | null | https://www.fullduplex.ai/benchmarks/msi-bench |
mmedfd | MMedFD | Academic (MMedFD authors) | Real-world healthcare benchmark for multi-turn full-duplex automatic speech recognition — adjacent to STS, crucial for duplex deployment. | true | adjacent | vertical | [
"asr",
"vertical-agent",
"full-duplex"
] | [
"duplex WER",
"medical-term recall",
"overlap-region WER"
] | 0 | 2025-09 | https://github.com/Kinetics-JOJO/MMedFD | https://arxiv.org/abs/2509.19817 | null | https://www.fullduplex.ai/benchmarks/mmedfd |
audio-jailbreaks-taxonomy | LALM Jailbreak Taxonomy & Eval | Feng et al. | Unified taxonomy of LALM jailbreak attacks (semantic, acoustic, signal, internal-representation) with controlled empirical evaluation under a single threat model and cost-aware protocol. | false | native | lab | [
"audio-safety",
"speech-lm"
] | [
"attack success rate",
"cost-aware ASR",
"defense utility",
"cross-attack generalisation"
] | 0 | 2026-05 | https://arxiv.org/abs/2605.30031 | https://arxiv.org/abs/2605.30031 | null | https://www.fullduplex.ai/benchmarks/audio-jailbreaks-taxonomy |
speechjbb | SpeechJBB | Ceccatelli, Jeon, Adelani et al. | Audio jailbreak benchmark for LALMs under multilingual and code-switched speech, with phonological perturbations as an augmented evaluation regime. | false | native | lab | [
"audio-safety",
"speech-lm"
] | [
"code-switched ASR",
"phonological-perturbation success",
"multilingual alignment gap"
] | 0 | 2026-08 | https://arxiv.org/abs/2606.06037 | https://arxiv.org/abs/2606.06037 | null | https://www.fullduplex.ai/benchmarks/speechjbb |
aia-acoustic-interference | AIA — Acoustic Interference Attack | Wang, Huang, Liang, Wu (CUHK · HKUST) | 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. | false | native | lab | [
"audio-safety",
"speech-lm"
] | [
"universal jailbreak success",
"alignment-evasion rate",
"transferability across LALMs"
] | 0 | 2026-05 | https://flaai.github.io/AIA_page/ | https://arxiv.org/abs/2605.18168 | null | https://www.fullduplex.ai/benchmarks/aia-acoustic-interference |
omni-duplexeval | Omni-DuplexEval | OpenBMB (Tsinghua University / Shanghai Qi Zhi Institute / ModelBest) | Streaming video benchmark that scores what a duplex omni-modal model says and when it says it, in live description and proactive reminder tasks. | true | adjacent | lab | [
"full-duplex",
"audio-understanding"
] | [
"overall score (0 to 100)",
"Content Consistency",
"Temporal Sensitivity",
"Proactive Reminder score"
] | 0 | 2026-07 | https://github.com/OpenBMB/Omni-DuplexEval | https://arxiv.org/abs/2605.17360 | null | https://www.fullduplex.ai/benchmarks/omni-duplexeval |
omniinteract | OmniInteract | CUHK MMLab / SJTU / NTU / McMaster / CityUHK / JUFE | Streaming audio-visual benchmark where spoken queries sit in the audio track and the model must decide whether, when and what to answer. | true | native | live | [
"full-duplex",
"audio-understanding"
] | [
"IA-QTF1 (Interaction-Aware Quality-Timeliness F1)",
"Interruption Diagnostic Suite (NOR, PAQ, CSM)",
"Nested Chain Completion Score"
] | 0 | 2026-05 | https://github.com/Lucky-Lance/OmniInteract | https://arxiv.org/abs/2605.26485 | null | https://www.fullduplex.ai/benchmarks/omniinteract |
provoice-bench | ProVoice-Bench | Shanghai Jiao Tong University (Xu et al.) | Tests whether a voice agent stays silent or steps in unprompted, across four proactive tasks built from synthesized conversations and device context. | false | native | lab | [
"speech-lm",
"audio-understanding"
] | [
"trigger accuracy",
"recall",
"false-positive rate",
"response accuracy (R_acc)"
] | 0 | 2026-06 | https://huggingface.co/datasets/overji/ProVoice-Bench | https://arxiv.org/abs/2604.15037 | null | https://www.fullduplex.ai/benchmarks/provoice-bench |
eot-bench | eot-bench | LiveKit | 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. | true | native | lab | [
"full-duplex"
] | [
"false cutoffs at 300 ms and 600 ms latency budgets",
"latency at 5% and 10% false-cutoff budgets",
"latency and false-cutoff Pareto frontier"
] | 0 | 2026-11 | https://github.com/livekit/eot-bench | null | https://livekit.com/benchmarks/eot-bench | https://www.fullduplex.ai/benchmarks/eot-bench |
audio2tool | Audio2Tool | Rivian and Volkswagen Group Technologies | Spoken tool-calling benchmark for smart car, smart home and wearable assistants, in eight tiers from direct commands to overlapping speakers. | false | native | lab | [
"speech-lm",
"vertical-agent"
] | [
"tool accuracy",
"exact match",
"parameter F1"
] | 0 | 2026-09 | https://audio2tool.github.io/ | https://arxiv.org/abs/2604.22821 | https://huggingface.co/spaces/RVtech/Audio2ToolLeaderboard | https://www.fullduplex.ai/benchmarks/audio2tool |
c3-benchmark | C3 Benchmark | Ma et al. | English and Chinese test of whether spoken dialogue models handle ambiguity, omission, coreference and multi-turn context in speech. | false | native | lab | [
"speech-lm"
] | [
"accuracy (LLM-judged)",
"per-phenomenon accuracy"
] | 0 | 2025-10 | https://step-out.github.io/C3-web/ | https://arxiv.org/abs/2507.22968 | https://step-out.github.io/C3-web/ | https://www.fullduplex.ai/benchmarks/c3-benchmark |
cekura-s2s-agent-bench | Cekura Speech-to-Speech Agent Benchmark | Cekura | Vendor-run leaderboard that swaps realtime speech models into one voice agent and scores 82 clinic and Medicare calls, three runs each. | false | native | vertical | [
"vertical-agent",
"speech-lm"
] | [
"pass³ (all three runs pass)",
"success rate",
"data accuracy",
"median response time"
] | 0 | 2026-11 | https://benchmarks.cekura.ai/speech-to-speech | null | https://benchmarks.cekura.ai/speech-to-speech | https://www.fullduplex.ai/benchmarks/cekura-s2s-agent-bench |
cekura-turn-bench | Cekura Turn-Taking Benchmark | Cekura | Scripted calls that test whether realtime speech models wait through pauses, reply at turn ends, stop on cut-ins and keep talking through backchannels. | true | native | live | [
"full-duplex"
] | [
"pause handling",
"turn taking",
"barge-in yield",
"backchannel continue rate",
"end-of-turn latency"
] | 0 | 2026-11 | https://benchmarks.cekura.ai/speech-to-speech/turn-bench | null | https://benchmarks.cekura.ai/speech-to-speech/turn-bench | https://www.fullduplex.ai/benchmarks/cekura-turn-bench |
codett | CoDeTT | Bairong (BRVoice team) | 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. | true | native | lab | [
"full-duplex"
] | [
"4-action accuracy (ACC)",
"Semantic Misalignment Rate (SMR)"
] | 0 | 2026-04 | https://huggingface.co/datasets/YingaoWang-casia/CoDeTT | https://arxiv.org/abs/2603.25434 | null | https://www.fullduplex.ai/benchmarks/codett |
contextdialog | ContextDialog | Seoul National University (Kim et al.) | Spoken questions about something said earlier in a conversation, testing whether voice interaction models recall and use past turns. | false | native | lab | [
"speech-lm"
] | [
"GPT Score (five-point)",
"word error rate"
] | 0 | 2025-05 | https://huggingface.co/datasets/ContextDialog/ContextDialog | https://arxiv.org/abs/2502.19759 | null | https://www.fullduplex.ai/benchmarks/contextdialog |
coval-benchmarks | Coval Benchmarks | Coval | 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. | false | component | live | [
"asr",
"tts",
"speech-lm"
] | [
"WER",
"TTFS",
"TTFA",
"voice-to-voice latency",
"instruction adherence"
] | 0 | 2026-11 | https://benchmarks.coval.ai/overview | null | https://benchmarks.coval.ai/overview | https://www.fullduplex.ai/benchmarks/coval-benchmarks |
duplexworld | DuplexWorld | Centific / University of Maryland | Speech-to-speech voice agents on simulated telephone tasks in six worlds, scored on task completion, conversational dynamics and speech naturalness. | true | native | vertical | [
"vertical-agent",
"full-duplex"
] | [
"Pass@1",
"goal-state check",
"turn-taking",
"selectivity",
"DNSMOS"
] | 0 | 2026-08 | https://huggingface.co/datasets/CentificAIResearch/Duplex-World | https://arxiv.org/abs/2608.10716 | null | https://www.fullduplex.ai/benchmarks/duplexworld |
game-time | Game-Time Benchmark | MIT / NTU / Academia Sinica | Tests whether full-duplex spoken language models can follow spoken instructions about timing, tempo and simultaneous speech. | true | native | lab | [
"full-duplex",
"instruction-following"
] | [
"instruction following (1 to 5)",
"response appropriateness (1 to 5)",
"turn-taking (1 to 5)"
] | 0 | 2026-06 | https://github.com/ga642381/Game-Time-Benchmark | https://arxiv.org/abs/2509.26388 | null | https://www.fullduplex.ai/benchmarks/game-time |
ihbench | IHBench | Boson AI | Scores what a voice agent says after the user interrupts it in the middle of a structured workflow. | false | native | lab | [
"speech-lm",
"vertical-agent"
] | [
"task fulfillment win rate",
"recovery quality pass rate"
] | 0 | 2026-06 | https://huggingface.co/datasets/bosonai/IHBench | https://arxiv.org/abs/2606.19595 | null | https://www.fullduplex.ai/benchmarks/ihbench |
kvoicebench | KVoiceBench | KRAFTON | Korean spoken question answering benchmark for speech language models, transferred from VoiceBench with human-agent translation and normalization rulebooks. | false | native | lab | [
"speech-lm",
"instruction-following"
] | [
"accuracy",
"judge score (open-ended)",
"prompt and instruction accuracy (KIFEval)",
"refusal rate (KAdvBench)"
] | 0 | 2026-08 | https://huggingface.co/datasets/KRAFTON/KVoiceBench | https://arxiv.org/abs/2605.27984 | null | https://www.fullduplex.ai/benchmarks/kvoicebench |
mivas-bench | MIVAS Bench | Bluejay | Speech-to-speech models placed inside multi-agent call environments for healthcare, legal services and customer support, scored by deterministic verifiers. | false | native | vertical | [
"vertical-agent",
"speech-lm"
] | [
"Pass^5",
"Pass@1",
"tool adherence",
"handoff adherence",
"final database state"
] | 0 | 2026-11 | https://github.com/bluejay-ai-dev/mivas-bench | null | https://research.getbluejay.ai/benchmarks/mivas/leaderboard | https://www.fullduplex.ai/benchmarks/mivas-bench |
openaudiobench | OpenAudioBench | Baichuan Inc. | Spoken question answering set from Baichuan covering logical reasoning, general knowledge and open-ended questions, judged by GPT-4o. | false | native | lab | [
"speech-lm"
] | [
"accuracy (Llama Questions, Web Questions, TriviaQA)",
"score (Reasoning QA, AlpacaEval)"
] | 0 | 2025-01 | https://huggingface.co/datasets/baichuan-inc/OpenAudioBench | https://arxiv.org/abs/2501.15368 | null | https://www.fullduplex.ai/benchmarks/openaudiobench |
rolebreak | RoleBreak | University of Hong Kong / Kami AI | Tests whether speech-to-speech models keep a persona, stay safe and express emotion over role-play conversations of about 22 turns. | false | native | lab | [
"speech-lm",
"emotion"
] | [
"persona pass rate",
"interaction pass rate",
"safety pass rate",
"first-failure turn (P-FFT, S-FFT)",
"naturalness",
"emotion"
] | 0 | 2026-09 | https://huggingface.co/datasets/Greenbean/RoleBreak | https://arxiv.org/abs/2609.16614 | null | https://www.fullduplex.ai/benchmarks/rolebreak |
s2s-arena | S2S-Arena | FreedomIntelligence / CUHK-SZ / Shenzhen University of Advanced Technology | Arena-style pairwise comparison of speech-to-speech models on spoken instructions that depend on paralinguistic cues, judged directly on audio. | false | native | lab | [
"speech-lm",
"instruction-following",
"emotion"
] | [
"Elo",
"win rate"
] | 0 | 2026-06 | https://github.com/FreedomIntelligence/S2S-Arena | https://arxiv.org/abs/2503.05085 | null | https://www.fullduplex.ai/benchmarks/s2s-arena |
speechinstructbench | SpeechInstructBench | CUHK / Alibaba Group | Spoken instruction-following benchmark in English and Chinese with closed-ended, open-ended and adjustment tasks under accent, noise and disfluency variants. | false | native | lab | [
"instruction-following",
"speech-lm"
] | [
"prompt-level accuracy",
"instruction-level accuracy",
"IAR",
"ECR"
] | 0 | 2026-01 | https://github.com/dingdongwang/SpeechInstructBench | https://arxiv.org/abs/2503.02769 | null | https://www.fullduplex.ai/benchmarks/speechinstructbench |
steerbench | SteerBench | Scale AI / University of Maryland | 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. | false | native | lab | [
"speech-lm",
"instruction-following"
] | [
"audio APR",
"sample APR",
"rubric pass rate"
] | 0 | 2026-11 | https://huggingface.co/datasets/utkarsh4430/SteerBench | https://arxiv.org/abs/2609.12623 | null | https://www.fullduplex.ai/benchmarks/steerbench |
televal | TELEVAL | China Telecom AI (Tele-AI) | 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. | false | native | lab | [
"speech-lm",
"audio-understanding",
"emotion"
] | [
"average score",
"Reliable Content Fulfillment",
"Interactional Appropriateness",
"speech-text consistency",
"DNSMOS"
] | 0 | 2026-08 | https://github.com/Tele-AI/TELEVAL | https://arxiv.org/abs/2507.18061 | null | https://www.fullduplex.ai/benchmarks/televal |
tpi-bench | TPI-Bench | Seoul National University / Yonsei University / University of Seoul | 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. | false | native | lab | [
"speech-lm",
"audio-understanding"
] | [
"Response Strategy Following (RSF)",
"Overall Helpfulness (OH)",
"BLEU",
"ROUGE-L"
] | 0 | 2026-04 | https://tpi-va.github.io/ | https://arxiv.org/abs/2604.17358 | null | https://www.fullduplex.ai/benchmarks/tpi-bench |
vera | VERA | Duke University / Adobe | Measures how much reasoning accuracy a model loses when the same questions are asked and answered by voice under real-time limits. | false | native | lab | [
"speech-lm"
] | [
"accuracy (macro-averaged over tracks)",
"time to first response (TTFR)",
"WER"
] | 0 | 2025-09 | https://github.com/linyueqian/VERA | https://arxiv.org/abs/2509.26542 | null | https://www.fullduplex.ai/benchmarks/vera |
voicebench-ja | VoiceBench-ja | SB Intuitions | Japanese test set for the gap between spoken and typed input, with the same prompts as audio and as text in four subsets. | false | native | lab | [
"speech-lm",
"instruction-following"
] | [
"LLM-judge score (1 to 5)",
"format following",
"exact match"
] | 0 | 2026-03 | https://huggingface.co/datasets/sbintuitions/voicebench-ja | https://arxiv.org/abs/2603.12565 | null | https://www.fullduplex.ai/benchmarks/voicebench-ja |
voxeval | VoxEval | CUHK / Lightspeed Studios / NUS | MMLU in spoken form: multiple-choice knowledge questions asked and answered in speech, repeated across speakers, speaking styles and audio quality. | false | native | lab | [
"speech-lm"
] | [
"accuracy (ASR + string matching)"
] | 0 | 2025-07 | https://github.com/dreamtheater123/VoxEval | https://arxiv.org/abs/2501.04962 | null | https://www.fullduplex.ai/benchmarks/voxeval |
wavbench | WavBench | Xiamen University / Zhejiang University / CUHK-Shenzhen | Scores spoken dialogue models on hard reasoning, on whether answers sound like natural talk, and on hearing and producing paralinguistic cues. | false | native | lab | [
"speech-lm",
"emotion",
"audio-understanding"
] | [
"colloquial score (Pro, Basic)",
"explicit understanding accuracy",
"explicit generation accuracy",
"implicit dialogue score"
] | 0 | 2026-02 | https://naruto-2024.github.io/wavbench.github.io/ | https://arxiv.org/abs/2602.12135 | null | https://www.fullduplex.ai/benchmarks/wavbench |
wearvox | WearVox | Meta | Voice assistant benchmark recorded on AI glasses, with 3,842 multi-channel egocentric clips that include real indoor and outdoor noise and bystander speech. | false | native | lab | [
"speech-lm",
"audio-understanding"
] | [
"turn-based micro-average accuracy",
"tool-call AST accuracy",
"side-talk rejection accuracy",
"speech translation judge score"
] | 0 | 2026-04 | https://github.com/facebookresearch/WearVox | https://arxiv.org/abs/2601.02391 | null | https://www.fullduplex.ai/benchmarks/wearvox |
wildspeech-bench | WildSpeech-Bench | Tencent (WeChat AI) | 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. | false | native | lab | [
"speech-lm"
] | [
"average score (1 to 10)",
"per-category score",
"UTMOS"
] | 0 | 2025-09 | https://github.com/Tencent/WildSpeech-Bench | https://arxiv.org/abs/2506.21875 | null | https://www.fullduplex.ai/benchmarks/wildspeech-bench |
Fullduplex catalogue
The models, datasets and benchmarks for full-duplex and speech-to-speech voice AI that fullduplex.ai tracks, one row per entry. Every row links to the entry's own pages (official page, paper, repository) and to its page on fullduplex.ai.
| Config | Rows | One row per |
|---|---|---|
models |
204 | model, hosted API, voice-agent platform or framework |
datasets |
126 | speech dataset |
benchmarks |
93 | benchmark |
Snapshot of 2026-11-16. The site is updated every week, so it can be newer than this file.
Fields worth knowing
full_duplex: for models, one model listens and speaks at the same time (duplex_typeisfull-duplex; the others arecascade,half-duplexandother). For datasets, each speaker is recorded on a separate channel. For benchmarks, the benchmark measures full-duplex behaviour such as turn-taking and interruptions.first_response_ms: the vendor's own claim, not a measurement by fullduplex.ai.speaker_consent:consented,public(collected from public media),mixedorunknown.license_class:permissive,share-alike,non-commercial,gatedorcustom. The exact terms are inlicense.page_url: the entry's page on fullduplex.ai, with the sources and the weekly reports that mention it.
Load it
from datasets import load_dataset
models = load_dataset("Fullduplex-ai/catalog", "models", split="train")
Licence and citation
CC BY 4.0. Please credit Fullduplex and link to https://www.fullduplex.ai.
@misc{fullduplex_catalog,
title = {Fullduplex catalogue: models, datasets and benchmarks for full-duplex voice AI},
author = {{Fullduplex}},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/Fullduplex-ai/catalog}}
}
- Downloads last month
- 173