cekura/turn-taking-bench

Cekura Turn-Taking Benchmark: public set (v1)

Can a voice AI hold a natural phone conversation? This set tests four behaviours that decide it:

• Waits through pauses. The caller pauses mid-thought ("my number is 415… 555…") and the AI should not cut in.
• Replies when the caller is done. The AI should answer once the caller has finished.
• Stops when the caller cuts in. The caller interrupts the AI ("no, not Tuesday, Thursday") and the AI should stop.
• Keeps going through "mm-hmm". The caller says "mm-hmm" or "okay" to show they are listening, and the AI should keep talking.

Leaderboard scores are computed on a private set built the same way (see "Public and private sets"). This public set lets anyone check the method and test their own system.

Contents

Calls	45 = 15 scripts × 3 caller/agent voice pairings, quiet-room audio
Behaviour types (scripts)	mid-sentence pauses (6), turn endings (1), long turns (1), cut-ins (5), "mm-hmm"s (2)
Caller voices	US female, US male, UK female (the voices every model was tested with)
Audio conditions	clean (quiet room)
Labels	every mid-turn pause, turn ending, cut-in and "mm-hmm", with exact timings

Files:

• public-00000-of-00001.parquet: one row per call. Caller audio (speaker_1_audio), the scripted agent side (speaker_2_audio), per-speaker event annotations, and authored metadata (script id, behaviour type, voices, condition).
• ground_truth.json: the planted events per call, including scripted vs detected pause lengths.
• durations.json: audio durations.
• manifest.json: file hashes; the folder loads directly as a turn-taking-bench dataset (--dataset <this folder>).
• specs/: the 15 script definitions (lines, pauses, cut-in and "mm-hmm" placements).
• calls_by_test.json: which calls feed which test.
• run-outputs/: our measured results on this set (see below).
• PRIVATE_SET_SHA256.txt: checksums of the private set's files, so anyone can later verify it never changed.
• interruption-clips/renders/: the exact caller clips (WAV + JSON sidecar) used to run the cut-in and "mm-hmm" test live, for the 3 caller voices. Pass this folder as --cache-dir to turn-taking-bench s2s t2-run so a re-run streams the same caller audio we did.

How the calls were made

Each script was voiced with ElevenLabs (eleven_turbo_v2_5) and assembled so every pause, cut-in and "mm-hmm" sits at a known time. Labels are verified with the end-of-turn scorer's gold builder. No real people's recordings are included; all voices are synthetic stock voices.

How results are scored

• End of turn (waiting through pauses, replying when done): Sesame TurnBench's published scorer, unchanged. Overall = mean of pause handling and turn taking.
• Cut-ins and "mm-hmm"s: each model's live reply is interrupted at fixed offsets; we measure whether and how quickly it stops, compared against matched moments with no interruption. This part is scored separately and never blended with end of turn.

Which calls feed which test

calls_by_test.json lists them. End of turn uses the 24 calls of the pause, turn-ending and long-turn scripts; cut-ins and "mm-hmm"s use the 7 cut-in and backchannel scripts, run live once per caller voice (21 sessions). Scoring end of turn on all 45 calls mixes in the interruption scripts and will not match our numbers.

turn-taking-bench run --dataset <downloaded folder> --model s2s:<provider> --ids $(jq -r '.end_of_turn.conversation_ids|join(",")' calls_by_test.json)
turn-taking-bench s2s t2-run --specs specs --cache-dir interruption-clips/renders --split dev --caller-voice f-us --provider <provider> --out runs/<model>-f-us

Our run outputs (public set, quiet room)

run-outputs/ holds what we measured on this set, so anyone can check our published public-set numbers exactly. No model audio is included.

• end_of_turn/<model>/: predictions, run record and report for each of the 13 models and 3 references; end_of_turn/dataset/ is the exact 24-call dataset those runs were pinned to. turn-taking-bench report --run <copy of a model folder> --dataset run-outputs/end_of_turn/dataset reproduces every number in its report.
• interruptions/sessions/<model>-<voice>/t2-events.json: every planned cut-in and "mm-hmm" with its outcome (stopped or kept talking, time to stop, or why it was dropped), plus the matched no-interruption moments; the per-session and combined summaries recompute exactly from these.

Reproducing results

• Rescoring saved predictions gives identical numbers: the end-of-turn scorer is pinned and confidence intervals use a fixed seed.
• Re-running live models lands within our 95% intervals, not identical: live models vary run to run, and the cut-in test is timed to each model's live reply.
• Location matters: reply and stop times include the network path. We ran from cloud servers in US East; a slow or busy machine drops more moments.
• Providers can update a model behind the same name; our results record the model id and run date.

Public and private sets

Set	Scripts	Use
Public (this dataset)	15	Released for anyone to run
Private	25	Leaderboard scores; never released
Reserve	10	Sealed; used for the next version

Scripts are assigned to sets per script, balanced by behaviour type, and every recording of a script stays in its set, so the public set does not reveal the private one. We report results on both sets; a model that does much better on the public set than on the private set suggests it was tuned to it.

Limitations

• English only; three synthetic caller voices (US and UK).
• Scripted customer-service calls (mostly appointment booking, account and card details, orders).
• Non-speech sounds (coughs, laughter) are not included yet.

License

CC BY 4.0. You may share and adapt this dataset, including commercially, with attribution to Cekura.