On trace T05, Rahul paused for 720 ms before saying a date, and Asha took the floor after 300 ms of silence and pushed for same-day payment over him. On trace T02, Sunita said "maine already pay kar diya" while Asha was stating the dues; Asha kept talking for 900 ms, then asked her for money anyway.
The first is an endpointing operating point; the second is barge-in handling and, separately, recovery. Each needs its own number.
What happened on the two calls?
In T05 the turn detector closed Rahul's turn during a hesitation before a date, and the agent's audio started over him 230 ms after he resumed. In T02 the barge-in monitor took 900 ms to stop the agent's speech, and the next agent turn ignored what Sunita had said.
T05, 16 November, TW-early-nov at its old 300 ms silence budget:
t_ms ch event / text
51200 caller vad_start REF: "main... agle mahine ki... [pause 720 ms] ...paanch tarikh ko kar dunga"
52950 caller pause begins after "agle mahine ki"
53250 agent eot_commit (silence 300 ms; TW-early-nov budget was 300 ms)
53670 caller pause ends (720 ms); REF: "...paanch tarikh ko kar dunga"
53900 agent TTS starts over the caller: "Kya aap aaj hi payment kar sakte hain? Late fee bach jaayegi."
Diagnosis: an early jump-in at 53,250 ms, then the agent's audio started over the caller at 53,900 ms, and the date was lost under it. Two defects sit in this trace: the detector committed on a hold pause, and playback then started 230 ms after the caller was speaking again. Fixing the second (cancel playback when caller speech resumes between eot_commit and play_start) turns some false cutoffs into a short delay instead of talk-over.
T02, 16 November, CD-short-nov, kitchen noise and a TV in the background:
t_ms ch event / text
30500 agent TTS (playing): "Sunita ji, aapke consumer loan ka EMI ek hazaar teen sau rupaye, jo paanch November ko due tha..."
32160 caller vad_start REF: "arre maine already pay kar diya, teen tarikh ko, PhonePe se"
32160 agent still playing (TV noise had raised the barge-in threshold)
33000 agent barge_in detected; tts_stop at 33060 (stop latency 900 ms)
34980 caller vad_end
35420 agent TTS: "Ji, samajh gayi. Toh kya aap aaj tak payment kar sakti hain?"
Diagnosis: a missed barge-in (900 ms of agent speech after the caller started, against a 500 ms limit) and a recovery failure. The stop latency is 840 ms of detection, slowed by the noise, plus 60 ms to tts_stop. The recovery failure is worse: no check_payment_status, no log_dispute, and a payment request after an "already paid" claim, a C08 failure. The span fields used here are defined in what a voice agent trace contains.
Why is endpointing a curve and not a number?
A turn detector trades two errors. Wait longer after the caller goes quiet and you cut off fewer callers who were only pausing, but every finished turn gets more dead air. Each detector gives a curve of false cutoffs against waiting time, and a point on it is a choice, not a property of the detector.
Endpointing is the decision that the caller has finished, so the agent may take the floor. In a cascaded pipeline it is the turn detector's eot_commit. Two quantities describe any setting:
- False-cutoff rate: the share of labelled hold pauses (the caller had not finished) on which the detector committed. Each such commit is a false cutoff.
- Dead air: on pauses labelled as true ends, the silence between speech end and the commit: the endpointing part of the end-of-turn delay, which measuring voice agent latency defines with the rest of the latency vocabulary.
People do not wait for silence. Across languages the most common gap between conversational turns is around 200 ms 1, and in a 10-language study of answers to yes/no questions the mean response offset was 208 ms and the median 100 ms, with overlaps counted as negative offsets 2. A detector that decides from silence alone cannot get near those gaps without cutting people off.
The best public benchmark of the trade-off is vendor-built. LiveKit's eot-bench (LiveKit made the benchmark, the dataset and the top-ranked model) scores end-of-turn detectors on the false-cutoff rate at a 300 ms and a 600 ms latency budget, and on the latency at a 5% and a 10% false-cutoff budget, where latency means dead air after the true end of turn, not compute time. Its dataset has 7,342 real human-to-agent user turns in 14 languages in a single validation split (Hindi 400 rows, English 2,288), and every silence of at least 100 ms in a turn is labelled eot (true end of turn) or hold (mid-turn pause). A dash in its tables means no setting reached that budget, not zero 3.
CAUTION
Vendor-reported. Treat it as a hypothesis to test on your own calls.
On pooled English (general, accented and short_turns cohorts, v1.1), eot-bench reports these false-cutoff rates at the two budgets, for four of the rows in its README 3:
| Detector | False cutoffs at 300 ms | At 600 ms |
|---|
| LiveKit Turn Detector v1.1 | 13.4% | 7.5% |
| Deepgram Flux | 22.3% | 13.5% |
| SmartTurn v3.2 | 46.8% | 22.9% |
| Silence-only VAD baseline | 79.3% | 33.4% |
The numbers move with the set. On LiveKit's June 2026 English set, Turn Detector v1 had a 9.9% false-cutoff rate at a 300 ms mean-latency budget, against 12.9% for Deepgram Flux and 27.7% for ultraVAD 4. On the harder v1.1 set (November 2026) the same column reads 13.4% for Turn Detector v1.1, 22.3% for Flux and 39.7% for ultraVAD 3: ultraVAD's rate rose 12 points between the two sets. This is LiveKit's own benchmark; rerun it on your audio before you quote it.
The second reason is the Hindi report. On 400 Hindi rows (vendor-run), the best false-cutoff rate at a 300 ms / 600 ms budget is 6.9% / 1.6% for LiveKit Turn Detector v1.1, 16.8% / 3.8% for v1.1-mini, 19.2% / 4.5% for Deepgram Flux, 42.6% / 6.5% for SmartTurn v3.2 and 49.5% / 7.4% for the silence-only VAD baseline, against 79.3% / 33.4% for the same baseline in English. Best mean latency at a 5% false-cutoff budget is 342 ms for LiveKit v1.1, 583 ms for Flux, 639 ms for SmartTurn v3.2 and 700 ms for the VAD baseline, with thresholds tuned separately for each budget 3.
Read the baseline row first. Silence alone cuts off 49.5% of Hindi holds and 79.3% of English holds at 300 ms, so the two sets differ in their pauses before any detector is involved; the Hindi table does not show that detectors are better at Hindi. The report does not say whether the Hindi rows include Hinglish code-mixing or 8 kHz telephone audio, and the dataset audio is distributed at 16 kHz, so these numbers do not describe a collections call 3. Keep eot-bench's four headline metrics and its labelling rule. Do not reuse its operating points.
Figure: Each detector is a curve; an operating point is a place on one curve. Synthetic Tarazu data, drawn to show shape only: detector B dominates A and silence-only at every budget, and the same detector on slower MFI-style pauses sits higher, which is why MFI-fort-nov runs a longer budget than TW-early-nov.
Which three failures should you count separately?
Count talk-over, early jump-in and missed barge-in as three metrics, because their causes and fixes differ. Talk-over is the agent speaking over a caller who held the floor; early jump-in is a false cutoff on a hold pause; missed barge-in is failing to stop, or to change course, when the caller interrupts.
Tarazu's definitions, computed from the trace on one clock:
- Talk-over: agent audio overlaps caller speech for more than 300 ms when the caller held the floor. For each agent playback interval and each caller speech segment that started before
play_start_ms, compute overlap = min(caller end, agent end) - max(caller start, play_start_ms). Rate: talk-over events per agent turn.
- Early jump-in (false cutoff): the agent takes the floor during a hold pause. On labelled pauses it is a commit on a pause labelled hold; on unlabelled production calls the proxy is caller speech that resumes after
eot_commit_ms and before, or just after, play_start_ms. Rate: false cutoffs over hold pauses, at the campaign's budget.
- Missed barge-in: the caller barges in with new information and the agent keeps speaking more than 500 ms, or does not revise its plan. Stop latency =
tts_stop minus the caller's vad_start for that barge-in. The first half is a stop-latency failure; the second is a recovery failure, scored on its own (see below).
Figure: The three turn-taking failures on one clock (grey = caller speech, accent = agent audio). Synthetic Tarazu data, with timings shaped like T05 (the middle lane) and T02 (the bottom lane). Each lane has a different cause: playback that does not yield, a detector that commits too soon, and a barge-in monitor that stops too late.
The fixes show why the split matters. Talk-over is cured by playback that yields when caller speech resumes. Early jump-in is cured by the operating point: a longer budget or a higher threshold in the right call states. Missed barge-in is cured by the barge-in monitor: faster detection, noise thresholds that do not drift up as they did in T02, and a stop path with no buffered audio to drain. A longer budget does nothing for T02; a faster monitor does nothing for T05. A single "interruption rate" moves when any of them moves and cannot say which. They also trade against each other: a monitor that stops instantly stops on every "haan haan".
Measure all three on the platform clock first, from spans, then check them at the handset on the dual-channel recording. The carrier legs shift the failures in opposite directions. After a barge-in, the caller keeps hearing the agent for the uplink plus downlink delay after the platform stops playback, so the stop latency the caller lives through is longer than the span says. For talk-over on a caller who held the floor, the same round trip makes the overlap at the handset shorter than on the platform. Recovering the carrier offset is covered in measuring voice agent latency. Report both clocks; the handset number is the one the caller lives through.
How do you build a Hinglish end-of-turn test set?
Take real calls, find every silence of at least 100 ms inside a caller turn, and label each end or hold, with a content tag. Use eot-bench's schema so the same four metrics apply, and oversample the pauses that cost the most: hesitations before amounts and dates.
Tarazu's gold-v7 holds 400 labelled pauses per language group, built this way:
- Source from production audio. Consented, redacted dual-channel recordings on the real phone path: 8 kHz, real carriers, real background noise. A pause set recorded on headsets at 16 kHz measures a different detector.
- Segment the caller channel only, so agent audio does not hide caller pauses, and keep every silence of at least 100 ms.
- Label with the future in view. A hold if the same speaker continued the same turn, an end if the floor passed. Record duration, call state (default, amount-asked, date-asked), language group and content tag (date, amount, small talk, dispute).
- Oversample the expensive pauses. Hesitations before amounts and dates ("agle mahine ki... paanch"), one-word replies ("haan", "ji", where a true end looks like a hold to a text model), side talk, and TV or traffic noise.
- Double-label a slice. Two bilingual reviewers label the same pauses, and you compute Cohen's kappa before anyone tunes a threshold. A label you cannot agree on cannot set a threshold.
- Keep the call id on every row. A hesitant speaker hesitates everywhere, so intervals resample whole calls.
Two Hinglish traps. A pause after "toh" or "but" at a language switch often means the caller is about to switch, not finish, and a text model trained on monolingual turns reads the clause as complete. And Hindi amounts are long words with natural breaks inside them ("chaar hazaar do sau pachaas"); a cutoff there splits the entity in two.
How do you choose an operating point per campaign and call state?
Price both errors in one unit, compute expected cost per 1,000 turns at each candidate setting, and pick the cheapest, separately for each campaign and each call state. A false cutoff before an amount or a date costs more than one in small talk, so states where Asha has just asked for a number get longer budgets.
Tarazu's cost model: one unit is one second of dead air on a finished turn. A false cutoff costs 1.0 units on small talk, 1.5 during a dispute and 2.5 before an amount or a date, because a cutoff there loses the entity the call exists to capture. The campaign owner signs these weights; they deserve more argument than the threshold.
A worked example on 40 labelled pauses from TW-early-nov, synthetic and made up for this page, at four silence budgets with the detector at thr 0.70. A hold pause is cut off if it lasts at least as long as the budget and the detector does not hold it on its own; dead air on a true end equals the budget, plus 300 ms on one-word replies the detector is unsure of.
state ends holds (ms, tag, H = detector holds at thr 0.70)
default 12 130 st, 180 st, 320 st, 370 st, 450 st, 560 st H, 340 dispute H, 610 dispute
(2 of the 12 ends are one-word replies: +300 ms dead air each)
amount-asked 8 140, 210, 280, 330, 360 H, 390, 440, 480, 550, 720 H, 760, 940 (all before an amount or date)
(1 of the 8 ends is a one-word reply)
Per state, the false-cutoff rate on holds and the cost:
| State | Budget | False cutoffs | Weighted cutoffs | Dead air (s) | Cost per 1,000 turns |
|---|
| Default (12 turns) | 300 ms | 4/8 = 50.0% | 4.5 | 4.2 | 725 |
| 400 ms | 2/8 = 25.0% | 2.5 | 5.4 | 658 |
| 500 ms | 1/8 = 12.5% | 1.5 | 6.6 | 675 |
| 600 ms | 1/8 = 12.5% | 1.5 | 7.8 | 775 |
| Amount-asked (8 turns) | 300 ms | 7/12 = 58.3% | 17.5 | 2.7 | 2,525 |
| 400 ms | 5/12 = 41.7% | 12.5 | 3.5 | 2,000 |
| 500 ms | 3/12 = 25.0% | 7.5 | 4.3 | 1,475 |
| 600 ms | 2/12 = 16.7% | 5.0 | 5.1 | 1,263 |
Cost per 1,000 turns is (weighted cutoffs + dead air) x 1,000 / turns; for the default state at 400 ms, (2.5 + 5.4) x 1,000 / 12 = 658. Pooled, false cutoffs fall from 11/20 = 55% at 300 ms to 15% at 600 ms while mean dead air rises from 345 to 645 ms.
What the table shows:
- The default state picks 400 ms. 300 -> 400 ms removes 2.0 units of cutoffs for 1.2 s of extra dead air; going on to 500 ms removes 1.0 unit for another 1.2 s, a net loss. That is the shape of the T05 fix on 17 November.
- The amount-asked state keeps buying until 600 ms, because each cutoff removed is worth 2.5 units and each 100 ms costs 0.8 s over 8 turns.
- A pooled curve picks one compromise, wrong for both states.
- Two amount holds are still cut at 600 ms (760 and 940 ms). The closed read-back catches those, not a longer budget.
- The set is oversampled. Most of its holds are 300 ms or longer by design; reweight each stratum to production frequencies before comparing with a tripwire, or the tripwire fires on your sampling plan.
Forty pauses cannot pin a choice. In the endpointing recipe's synthetic run on 600 pauses, TW-early-nov's default-state choice moves between 250 and 500 ms across call-clustered resamples, so the team's 400 ms is consistent with the data rather than determined by it. If a plausible change in the weights moves a campaign by 200 ms, the weights are the decision.
Tarazu's operating points (from 2026-11-17):
| Campaign | Default silence budget | Amount-asked / date-asked | Detector threshold |
|---|
| TW-early-nov | 300 ms (before T05) -> 400 ms | 300 ms (before T05) -> 600 ms | td-1.1 thr 0.62 -> 0.70 |
| PL-digital-nov | 400 ms | 600 ms | thr 0.70 |
| CD-short-nov | 400 ms | 600 ms | thr 0.70 |
| SBL-guar-nov | 450 ms | 650 ms | thr 0.72 |
| MFI-fort-nov | 600 ms | 800 ms | thr 0.78 |
MFI-fort-nov gets the longest budgets because its older callers pause longer, so its curve sits higher (the dashed line in the first figure). SBL runs slightly longer because its amounts are said in lakh.
Two rules keep these numbers honest. First, the operating point belongs to a detector version, so the trace pins it (td-1.1@thr0.62 in the trace schema) and every reported rate names it. Second, rerun the curve whenever the detector, the STT model or the agent SDK changes: td-1.1 reads STT partials, so an STT upgrade moves its curve, and an SDK upgrade can change a default you never chose. The general problem is covered in environmental drift.
How do you tell a barge-in from a backchannel?
Label every caller sound during agent speech as barge-in, backchannel, side talk or background, then score the agent's stop decision against those labels. A barge-in (new information or a request to stop) should stop the agent; a backchannel ("haan haan", "accha", "hmm"), side talk and background noise should not.
Side talk is speech to someone else in the room; background is TV or traffic. Score it as a four-row confusion table: for each label, the share of events on which the agent stopped. Report stop latency on the barge-in row only, as a P95 sliced by noise condition, because T02's 840 ms detection delay came from a threshold the TV had pushed up.
Full-Duplex-Bench v1.5 shows why both rows matter. It tests five full-duplex systems on four overlap events (user interruption, user backchannel, user talking to others, background speech) using 200 synthesized interruptions. GPT-4o Realtime stops fastest on real interruptions (stop latency 0.23 s, respond rate 0.78) but also responds to background speech 93% of the time and to speech aimed at someone else 91% of the time, while Gemini and Nova Sonic hold the floor on side speech but take more than 2 s to stop on real interruptions. The authors find the same trade-off across systems: fast stopping comes with poor floor control on speech that was not meant for the agent 5. These are mid-2025 endpoints, English only, with simulated overlaps; current products may differ. The structure holds: a monitor tuned on the barge-in row alone fails the side-talk and background rows.
A collections call adds a rule that no benchmark scores. A backchannel is never a confirmation of a changed amount. In T01, Rahul's "haan haan theek hai" after Asha stated a wrong amount signalled listening, not agreement to Rs 2,000. During a read-back, "haan haan" while Asha is still reading is a backchannel; the explicit yes that C12 requires comes after the closed question ends. So the label feeds two decisions, whether to stop talking and whether the words count as an answer, and the second is gated on timing: words that overlap the question do not answer it.
Polarity inside an overlap matters too. "Nahi nahi" during a read-back is a barge-in, not a backchannel, even though it is two short repeated words like "haan haan". A classifier that learns "short and repeated means backchannel" lets a refusal pass as attention.
Why is recovery its own axis?
Stopping on time is half of handling a barge-in. The other half is changing course: the next agent turn must reflect what the caller said, and the right tool must run. In T02 a 500 ms stop would still have been followed by a payment request after an "already paid" claim, so recovery gets its own pass rate.
Current systems fail here. In the EchoChain benchmark (Modi et al., arXiv preprint, April 2026), four real-time voice models (GPT-realtime-2025-08-28, Gemini Live 2.5 flash native audio, Grok Voice Agent, Amazon Nova Sonic 2) had to revise their answer after a scripted user barge-in mid-response. Across 200 interrupted conversations none fully passed more than half, with Grok Voice Agent best at 48.5% and Gemini Live lowest at 16.5% 6. Those are speech-to-speech systems on scripted interruptions, not Asha, but they show that stopping and revising are different skills.
Tarazu's recovery check runs on every scenario card with an interruption script (content-triggered, not timer-based) and on flagged production barge-ins. A barge-in passes when all four hold:
- History truncation. The conversation history the LLM sees contains only the words that played before
tts_stop, not the full planned sentence. Otherwise the agent believes Sunita heard the whole dues statement, and later turns build on it.
- No resumption of the interrupted plan. The next agent turn does not repeat or finish the interrupted content unless the caller asks for it.
- The new information is reflected. The next agent turn acknowledges the claim in substance ("main abhi check karti hoon"), not with a filler followed by the old plan ("Ji, samajh gayi. Toh kya aap aaj tak payment kar sakti hain?").
- The required tool runs, in order. For "already paid":
check_payment_status before any payment request, then end_call(paid) if it finds the payment, or log_dispute(already_paid) if it does not, with the grievance details stated and send_grievance_sms before the call ends either way (the F02 sequence, policy C08).
Items 3 and 4 are trajectory checks, scored on what the agent did after the event; why a final-state check misses this is in trajectory vs outcome; the CRM diff for F02 is owned by task success and outbound call outcomes. How to generate callers who interrupt on content, with noise, is in simulating callers.
Report stop latency, recovery and their joint pass. A release that halves stop latency with recovery flat has improved the sound of the call, not its outcome.
Do full-duplex models fix turn-taking?
No. Full-duplex models listen and speak at once, which removes the cascade's fixed silence budget, but published benchmarks show the same trade-off in a new form: the systems that answer fastest also take the floor during the caller's pauses most often. Report reply latency only together with the takeover rate.
Full-Duplex-Bench measures turn-taking latency from the end of the user's speech to the start of the model's reply, counted only when the model takes the turn. On the Candor set Moshi answers in 0.265 s and dGSLM in 0.352 s, against 0.953 s for Freeze-Omni and 1.301 s for Gemini Live. But Moshi also takes over during user pauses almost every time (takeover rate 0.985 synthetic, 0.980 Candor), while Gemini Live takes over far less (0.255 and 0.310) 7. Low latency alone rewards a model that talks over pauses. Those are 2024-2025 system versions on English.
IndicFDB (NVIDIA and VoiceArena, September 2026) extends Full-Duplex-Bench to ten languages spoken in India, with 12,350 samples mined from about 50,000 hours of channel-separated conversations. In Hindi the human reference turn-taking latency is 0.259 s; GPT Live answers in 0.181 s but succeeds on only 61.59% of pause-handling and 15.32% of backchannel samples, while Gemini 3.8 Live handles 100% of pauses but answers after 1.400 s. The authors write that no evaluated system escapes this trade-off 8. The 0.259 s is the other speaker's response in mined conversations, not a phone-call norm, and the paper does not test code-mixed speech, most of Tarazu's traffic.
A full-duplex model moves the operating point from a config file into model weights: you can no longer set 600 ms for the amount-asked state, only measure what the model does there. The pause set, the overlap labels and the recovery suite become the only way to see the trade-off. What else changes if Asha's cascade is replaced by the asha-s2s-pilot speech-to-speech model, including which spans disappear, is covered in designing the voice evaluation platform.
What gate would have caught them?
Three release gates, each per language group and call state: the false-cutoff rate at the campaign's budget on the labelled pause set, the stop-latency P95 on the barge-in suite, and the recovery pass rate on the same suite. Each is reported with an interval and compared with the incumbent on the same items.
gate: turn-taking
gold_set: gold-v7 # 400 labelled pauses per language group; scenario cards with interruption scripts
detector: td-1.1 # threshold and budgets from the campaign table, pinned in the trace
segments: [hinglish, hindi-dominant, indian-english]
states: [default, amount-asked, date-asked]
metrics:
false_cutoff_rate:
at: campaign_budget # per campaign and call state
weights: {small_talk: 1.0, dispute: 1.5, amount: 2.5, date: 2.5}
interval: call-clustered bootstrap, 95%
stop_latency_p95:
events: barge-in labels only
slices: [quiet, tv, street]
interval: call-clustered bootstrap, 95%
limit_ms: 500
recovery_pass_rate:
checks: [history_truncated, no_resume, info_reflected, required_tool_in_order]
interval: wilson # one scenario, one pass or fail
compare: mcnemar # candidate vs incumbent on the same scenarios
report_always: [talk_over_rate, early_jump_in_rate, stop_rate_by_label, dead_air_mean]
block_if:
- false_cutoff_upper_bound > campaign_limit
- stop_latency_p95_upper_bound > 500
- recovery_pass_rate_lower_bound < campaign_floor
- paired_difference_worse_than_margin
The interval rules follow from what is counted. Pauses within a call are correlated, so the false-cutoff rate gets a bootstrap that resamples whole calls; Wilson over pauses would be too narrow. Stop latency is a percentile, so it is call-clustered too. Recovery is one pass or fail per scenario, so a segment's pass rate gets a Wilson interval, and candidate against incumbent is McNemar's test on the same scenarios. Expect wide intervals at nightly sizes: with 20 scenarios in a segment, even 20 of 20 passing gives a Wilson 95% lower bound of 0.839. Pass rates over repeated runs, and how many runs a gate needs, are in reliability statistics for voice agents.
Both calls fail these gates. T05's pattern, a long hold before a date, is the slice the pause set oversamples; at TW's old 300 ms budget the amount-asked and date-asked states fail the false-cutoff gate before release, and the talk-over check points at the missing cancel-on-resume. T02 fails twice on a suite that includes a TV-noise slice: stop latency of 900 ms against the 500 ms limit, and recovery item 4, because check_payment_status never ran. In production, the F02 tool-sequence check on 100% of calls fails T02 on the spot. Targets for the rest of a voice agent's metrics sit in the voice task-type playbook.
What goes in the decision memo?
The memo records one endpointing operating point per campaign and call state, the cost weights that chose it and who signed them, and the tripwire that reopens it. Below is Tarazu's memo after T05.
Decision. TW-early-nov moves from 300 ms at td-1.1 thr 0.62 to 400 ms at thr 0.70 in the default state, and 600 ms in the amount-asked and date-asked states. PL and CD run 400 / 600 ms at thr 0.70, SBL 450 / 650 ms at thr 0.72, MFI 600 / 800 ms at thr 0.78. Playback cancels when caller speech resumes before play_start.
Context. T05 (16 November): a false cutoff on a 720 ms hold before a date. TW-early-nov carries the most volume.
Options considered. (1) One global budget: wrong for MFI and for the amount-asked state. (2) Per campaign only: a compromise wrong for both states. (3) Per campaign and call state: chosen. (4) A full-duplex model: no evidence on code-mixed speech.
Evidence. gold-v7 labelled pauses (400 per language group), cost per 1,000 turns with weights 1.0 / 1.5 / 2.5, call-clustered bootstrap intervals; the recipe's synthetic run puts TW's default choice between 250 and 500 ms.
Owner and cadence. Turn-taking owner on the eval platform team; each campaign owner signs the weights. Rerun on every detector, STT or SDK version.
Tripwires. Hindi-dominant false-cutoff above 4% at the chosen operating point for two weeks; stop-latency P95 above 500 ms on any noise slice.
Dated facts. eot-bench README and Hindi report (19 November 2026); Full-Duplex-Bench v1.5 (April 2026); IndicFDB (September 2026); EchoChain (April 2026).
What to do this week
Label a small pause set from your own calls, draw your detector's curve on it, split your interruption metric into three, and test recovery separately from stop latency. Five steps:
- Pull 30 calls and label every caller pause of at least 100 ms as end or hold, with call state and a content tag; keep the call id on each row.
- Plot false cutoffs against dead air for your current detector on those pauses, and mark where your production setting sits.
- Split your "interruptions" dashboard into talk-over, early jump-in and missed barge-in, using the formulas above on your trace fields.
- Write five interruption scenarios where the caller says something that must change the plan ("already paid", "wrong number", "I lost my job"), and score stop latency and recovery separately.
- Check that your agent's conversation history is truncated to the words that actually played before a barge-in.
NOTE
What will date on this page. eot-bench figures are from its README and Hindi report as of 19 November 2026 (v1.1 released 13 November 2026); its tables change as detectors are added, so recheck the rows and cite the commit date. The June 2026 English figures come from LiveKit's blog. Full-Duplex-Bench v1.5 is arXiv v4 (April 2026) and tested mid-2025 endpoints; Full-Duplex-Bench is v3 (August 2025). IndicFDB is v1 (September 2026). EchoChain is an April 2026 preprint. The worked example's pauses are synthetic and made up for this page.