Full-Duplex-Bench v1: speech benchmark by NTU / UC Berkeley / UW / MIT

Benchmarks

verified 2026-11. Download on Hugging Face

TierSetting27 of 64

	Full-duplex	Tier	Setting	Focus
DuplexActDong et al.	Full-duplex	Speech-native	Lab	—	8	2026-10
Full-Duplex-Bench v1NTU / UC Berkeley / UW / MIT	Full-duplex	Speech-native	Lab	—	5	2025-03
VideoFDBNVIDIA / David AI	Full-duplex	Speech-native	Live	—	4	2026-10
TurnBenchSesame AI Labs / Mundo AI / CMU / oto	Full-duplex	Speech-native	Lab	—	3	2026-11
Full-Duplex-Bench v1.5NTU / UC Berkeley / UW / MIT	Full-duplex	Speech-native	Lab	—	3	2025-08
Talking TurnsApple / CMU	Full-duplex	Speech-native	Lab	Speech-LM	3	2025-04
τ-bench 2.0 (τ-Voice)Sierra Research	Full-duplex	Speech-native	Vertical	Vertical agent Speech-LM	3	2026-10
Full-Duplex-Bench v3NTU (NVIDIA advisory)	Full-duplex	Speech-native	Live	Vertical agent	2	2026-05
MultiPartyTurnBenchMathur and Manocha	Full-duplex	Speech-native	Lab	—	2	2026-10
Audio MultiChallengeScale AI (Scale Labs)	Full-duplex	Speech-native	Lab	Speech-LM	2	2026-10
Full-Duplex-Bench v2NTU / UC Berkeley / UW / MIT	Full-duplex	Speech-native	Live	Speech-LM	0	2026-02
MultiTalkBenchWang et al. (MultiTalk)	Full-duplex	Speech-native	Lab	—	0	2026-09
M3-DuplexBenchNTT (Fukuda et al.)	Full-duplex	Speech-native	Lab	Speech-LM	0	2026-10
Duplex-MPEMa et al.	Full-duplex	Speech-native	Lab	—	0	2026-11
DuplexLedgerLi et al.	Full-duplex	Speech-native	Lab	Speech-LM	0	2026-10
InteractionBenchAvera Labs (He et al.)	Full-duplex	Speech-native	Live	Speech-LM	0	2026-11
ECHOZhao et al.	Full-duplex	Speech-native	Lab	—	0	2026-09
MP-BenchShih et al.	Full-duplex	Speech-native	Lab	Speech-LM	0	2026-10
DuplexSpeechBench-IFEvalMathur and Manocha	Full-duplex	Speech-native	Lab	Instruction following	0	2026-09
APEX-VoiceMathur and Manocha	Full-duplex	Speech-native	Vertical	Vertical agent	0	2026-09
SID-BenchXia et al.	Full-duplex	Speech-native	Lab	—	0	2026-03
FD-BenchPeng et al.	Full-duplex	Speech-native	Lab	—	0	2025-07
MTR-DuplexBenchZhang et al.	Full-duplex	Speech-native	Lab	Speech-LM	0	2026-04
HumDial (ICASSP '26)ICASSP 2026 Spoken Dialogue Challenge	Full-duplex	Speech-native	Lab	Speech-LM Emotion	0	2026-01
FLEXIGe et al.	Full-duplex	Speech-native	Lab	—	0	2025-09
J-Moshi (subjective)Ohashi et al.	Full-duplex	Speech-native	Lab	Speech-LM	0	2025-06
MMedFDAcademic (MMedFD authors)	Full-duplex	Adjacent	Vertical	ASR Vertical agent	0	2025-09

Model

Pause handlinglower is better

Smooth turn-taking

Turn-taking latencylower is better

Reply after an interruption

Interruption latencylower is better

Backchannel timinglower is better

dGSLM

93.4%

97.5%

0.35 s

0.20

2.53 s

0.934

Moshi

98.5%

94.1%

0.27 s

0.77

0.26 s

0.957

Freeze-Omni

64.2%

33.6%

0.95 s

3.62

1.41 s

0.997

PersonaPlexpre-release model

58.4%

99.2%

0.07 s

4.21

0.40 s

0.649

Gemini 2.0 Flash Live

25.5%

65.5%

1.30 s

3.38

1.18 s

0.896

GPT-Realtimeversion not stated

1.0%

100.0%

1.47 s

3.85

1.50 s

0.980

Gemini 3.1 Flash Liveminimal thinking

2.2%

100.0%

0.57 s

3.58

0.34 s

0.807

Gemini 3.1 Flash Livehigh thinking

0.7%

99.2%

0.94 s

4.12

0.65 s

0.853

Model

Pause handlinglower is better

Smooth turn-taking

Turn-taking latencylower is better

Reply after an interruption

Interruption latencylower is better

Backchannel timinglower is better

Moshi

44.5%

73.9%

0.16 s

3.44

1.38 s

0.824

Moshi + RLKyutai's RL checkpoint

30.7%

95.8%

0.16 s

3.63

0.41 s

0.794

PersonaPlexreleased checkpoint

48.2%

95.8%

0.22 s

4.50

0.27 s

0.841

PersonaPlex + RLKyutai's RL checkpoint

35.0%

97.5%

0.09 s

4.53

0.22 s

0.786

Model

Pause handlinglower is better

Smooth turn-taking

Turn-taking latencylower is better

Reply after an interruption

Interruption latencylower is better

Backchannel timinglower is better

Raon-SpeechChat-9Bmodel authors' own model

21.2%

83.2%

1.03 s

2.79

1.22 s

0.775

MiniCPM-o 4.5

18.2%

89.1%

1.00 s

3.41

1.23 s

0.899

PersonaPlex

21.2%

78.2%

1.10 s

2.94

0.98 s

0.735

Moshi

29.9%

43.7%

0.73 s

2.91

2.03 s

0.664

Freeze-Omni

62.0%

25.2%

1.14 s

2.83

2.27 s

0.983

Model

Pause handlinglower is better

Smooth turn-taking

Turn-taking latencylower is better

Reply after an interruption

Interruption latencylower is better

Backchannel timinglower is better

PersonaPlexPersonaPlex paper, Table 6

35.8%

90.8%

0.17 s

4.29

0.24 s

0.662

NemotronLabs VoiceChat 11Bno backchannel track

15.3%

81.5%

0.45 s

4.33

0.48 s

—

MoshiRAGMoshiRAG paper, Table 2

32.0%

83.0%

0.18 s

3.75

1.02 s

0.940

Moshi + RLKyutai paper, Table 1

30.7%

95.8%

0.16 s

3.63

0.41 s

0.794

PersonaPlex + RLKyutai paper, Table 1

35.0%

97.5%

0.09 s

4.53

0.22 s

0.786

DuplexCascadeDuplexCascade paper, Table 1

5.8%

83.2%

1.72 s

4.02

1.23 s

0.949

DuplexCascade-βwith system backchannels

34.3%

89.9%

0.57 s

4.01

0.85 s

0.811