Which voice AI is best at turn-taking?

Averaged over all six line conditions, Amazon Nova 2 Sonic ranks first on end of turn at 0.85. Its 95% range overlaps the next 9, so Amazon Nova 2 Sonic, GPT Realtime 2.1, GPT-Live 1, Gemini 3.8 Live Ext. Thinking, Phonic v1, Gemini 3.1 Flash Live, Gemini 3.8 Live, Deepslate Opal, Azure Realtime and GPT Realtime 2.1 Mini are too close to separate. Cut-in handling is shown beside the rank and is not part of it.

Why do most models stop talking at filler words?

They treat any sound from the caller as an interruption. That makes them quick to stop when the caller really cuts in, but it means a simple “mm-hmm” halts them mid-sentence. Only GPT-Live 1 keeps going most of the time, and it pays for that by ignoring many real interruptions.

Why isn't cut-in handling part of the rank?

The rank uses Sesame TurnBench's published end-of-turn scorer. Cut-in handling is the mean of stopping on a real cut-in and carrying on through filler words, with our own scoring, so it sits beside the rank instead of inside it. Where fewer than 80% of moments could be scored it is marked indicative.

Were all models tested the same way?

Yes. Every model heard the same recorded calls, streamed from the same cloud region, with the provider's default turn detection. Reply and stop times include the network path from US East.

How does this compare with Artificial Analysis?

On the seven models both benchmarks test, the rankings largely agree (0.89). The one big difference is Gemini 3.1 Flash Live, which ranks near the bottom there and top here in a quiet room.

What is the public set?

15 public scripts, 45 calls in a quiet room with 3 US and UK caller voices, to be published at launch under CC BY 4.0 so anyone can rerun the method. The leaderboard ranks the 25-script private set, kept private so results cannot be tuned to it. Every model scores lower on the public set, and the order broadly agrees.

Can I test my own voice agent like this?

Yes. Cekura runs the same kind of scripted calls against your agent, under the same line conditions.

Bench

Turn-taking benchmark

End of turnHow we tested

Best for

Gemini 3.1 Flash Live
- Quiet rooms
- Phone calls
Amazon Nova 2 Sonic
- Quiet rooms
- Weak signal
GPT Realtime 2.1
- Loud noise
- Echoey rooms
GPT-Live 1
- Light noise
- Mid-sentence pauses
- Filler words
Phonic v1
- Phone calls

End of turn

1Amazon Nova 2 Sonic85
1GPT Realtime 2.185
3GPT-Live 184

Handles cut-ins

1Grok Voice 3.071
2Grok Voice Think Fast 2.064
3Azure Realtime58

Last updated Oct 13, 2026Methodology by Prashansa Dadu

What we found

Gemini 3.1 Flash Live leads in a quiet room at 88.4%. Noise is where models come apart.

The typical model, out of 13

Waits through a pause73%of pauses

Replies when you're done99%of finished turns

Stops when you cut in100%of cut-ins

Keeps going through filler words6%of filler words

Heavy noise

The quiet-room leader falls apart in noise

Gemini 3.1 Flash Live, finished turns answered

Quiet room97%Heavy noise55%

End of turn

Does the model wait while the caller is still talking, and reply once they finish?


Amazon Nova 2 Sonic	88	87	84	82	83	89	85
GPT Realtime 2.1	85	84	83	84	88	85	85
GPT-Live 1	86	81	85	81	86	87	84
Gemini 3.8 Live Ext. Thinking	83	86	81	82	86	85	84
Phonic v1	86	88	80	76	85	85	83
Gemini 3.1 Flash Live	88	88	79	72	85	85	83
Gemini 3.8 Live	85	85	77	78	83	82	82
Deepslate Opal	80	79	76	75	85	76	78
Azure Realtime	78	78	76	77	79	78	78
GPT Realtime 2.1 Mini	77	78	74	75	81	78	77
Grok Voice 3.0	82	74	67	60	82	81	74
Grok Voice Think Fast 2.0	82	78	67	57	81	78	74
Higgs Realtime	71	69	70	71	72	70	70

5789Yellow is weaker, green is stronger · out of 100 · click a heading to sort

Cut-ins and filler words

Most models stop when you cut in, but they also stop for a quick “mm-hmm”.

Quiet room

No processing. The caller sounds as clear as they can.

01Grok Voice 3.0*845871
02Grok Voice Think Fast 2.0*894266
03GPT-Live 1*347856
04Azure Realtime1001256
05Higgs Realtime100954
06Gemini 3.8 Live100753
07Gemini 3.8 Live Ext. Thinking*100653
08Gemini 3.1 Flash Live100553
09Phonic v1*98552
10Deepslate Opal100352
11Amazon Nova 2 Sonic100050
12GPT Realtime 2.1100050
13GPT Realtime 2.1 Mini100050

Out of 100. Stops: stops when the caller cuts in. Filler words: keeps talking through a quick “mm-hmm”. * Fewer than 80% of moments could be scored, so treat as rough. Click a heading to sort.

Results by call

Most models reply on time. Pauses are harder, most of all when the caller goes quiet to look something up.

Says “uh…” mid-sentence

The caller pauses with fillers like “uh”, “let me think” or “you know” in the middle of a sentence.

#ModelScore

01Gemini 3.8 Live Ext. ThinkingGoogle90%
02Gemini 3.1 Flash LiveGoogle83%
03Amazon Nova 2 SonicAmazon83%
04Phonic v1Phonic83%
05GPT-Live 1OpenAI80%
06GPT Realtime 2.1OpenAI80%
07Gemini 3.8 LiveGoogle80%
08Grok Voice Think Fast 2.0xAI80%
09Deepslate OpalDeepslate80%
10Azure RealtimeMicrosoft80%
11GPT Realtime 2.1 MiniOpenAI63%

Turn-taking score for each kind of call. Quiet room, earlier 12 Oct run, three calls per kind, so small gaps are noise.

Bad lines

The same calls replayed on a phone line, in noise, in an echoey room and with dropouts.

Gemini 3.1 Flash LiveAmazon Nova 2 Sonic

How we tested

The same scripts, played to every model. We timed four moments where the AI has to choose: talk or stay quiet.

1 · Pause

Waits through a pause

I need to move my appointment tonext Tuesday.AI stays quiet

2 · End of turn

Replies when you're done

…next Tuesday.Sure, Tuesday at 10 am works.

3 · Barge-in

Stops when you cut in

I'll also text you a remin…Wait, can it be Wednesday?

4 · Backchannel

Keeps going through filler words

Please bring your insurance card,mm-hmmand arrive ten minutes early.

13
models: 25
scripts: 6
line conditions: 36
calls per model, per condition

Benchmark code

Full methodology

One call, step by step

1
Pick the calls
Public and private sets
Scripts are assigned to a public set (15, released under CC BY 4.0) and a private set (25 leaderboard scripts plus 10 sealed reserve scripts). Leaderboard scores come from the private set; public-set scores are shown alongside so anyone can check the method. Only tested material is published: 3 US/UK caller voices, quiet-room audio for the public set.
2
Play them over six lines
Line conditions
Each call is played clean and through five degraded lines. Press play to hear the caller in each one.
3
Stream them live
Live streaming and the clock
Caller audio is sent in 20 ms chunks paced against a fixed origin; model audio is timestamped on arrival. All times are on one host clock and include the network path from US East to each provider and back.
4
Cut in while the model talks
Scheduling cut-ins and backchannels
Events are scheduled from the start of the model's live reply (the anchor): cut-ins at 1, 2 or 3 s into the reply; backchannels from 2.5 s, every 2.5 s, up to 4 per reply. At the target time the model must be speaking (Silero VAD over the last 0.6 s, polled every 20 ms, up to 0.5 s grace); then the caller clip is injected and its actual send time recorded.
5
Decide if it stopped
Did it stop, and when?
After injection the model's audio is observed for 1.2 s. Speech cessation is decided by Silero VAD with a residual-energy floor (+6 dB margin) and needs 0.2 s free of speech. Barge-in yield = stopped within 1.0 s of the cut-in. Stop latency = cut-in start to speech end (provisional). Backchannel continue rate = share of backchannels the model keeps talking through.
6
Compare with silence
No-interruption baseline
At matched offsets (1, 2, 2.5, 3, 5, 7.5 and 10 s) in replies with no stimulus, we measure how often the model stops anyway. Excess yield = yield minus this base rate.
7
Score pauses and turn ends
End of turn
Pause handling and turn taking are scored with Sesame TurnBench's published scorer, unchanged (pinned version). Overall = mean of the two. 95% intervals by whole-conversation bootstrap (2,000 draws, fixed seed). A model is ranked when at least 90% of its calls are scored; references are never ranked. Reply time is the median from the end of caller speech to the model's first audio on answered turns.
8
Drop unclear moments
Dropped moments and coverage
A moment is voided, never guessed, when the model gave no reply, was not speaking at the target, the target had passed, the audio was too unclear to judge, or our host fell behind: a send more than 25 ms late voids events whose window (1 s before the cut-in to the end of observation) overlaps it; a stall over 250 ms voids the session. Coverage = scored / planned moments; below 80% is shown as indicative.
9
Check the harness
Calibration
Our harness reproduces Sesame TurnBench's reference timing on its own data: 99.3% of reference caller events matched within 0.5 s, median offset 1 ms. Two separate cloud runs agreed on 20 of 23 model rows within their 95% intervals.

Can I reproduce it?

Rescoring our published outputs (run-outputs/ in the public dataset) gives our exact numbers: the scorer is pinned and intervals use a fixed seed.
Re-running live models gives numbers within the 95% intervals, not identical: live models vary run to run.
Location matters: reply and stop times include the network path; we ran from US East. A slow or busy machine drops more moments.
Providers can update a model behind the same name; results carry the model id and run date.
The private-set leaderboard cannot be re-run outside Cekura by design; public-set scores are the comparable ones.

What the terms mean

Waits through pauses pause handling: Share of mid-turn pauses where the model stays silent.
Replies when the caller is done turn taking: Share of real turn endings the model answers.
Reply time end-of-turn latency, median: End of caller speech to the model's first audio.
Stops when the caller cuts in barge-in yield: Stopped within 1 s of a cut-in.
Time to stop stop latency, median: Cut-in start to the end of the model's speech.
Keeps going through "mm-hmm" backchannel continue rate: Share of backchannels the model talks through.
Moments scored coverage: Scored over planned moments; under 80% is indicative.