Autonomous multi-angle capture

How often does a short window of sound sync real phones and cameras to within a video frame, and can a confidence set in one room hold in another?

research/capture-sync

Research programme: audio sync of free-running cameras and phones on real devices; step 1 on five real smartphones in a 200 ms and an 800 ms room (LibriWASN): clocks differ by up to 19 ppm, and with -10 dB of white noise GCC-PHAT misses a frame on 17-34% of 2 s windows against 5-7% for partial whitening…

52 tests

What it found

An open research programme (track 13, capture) on syncing free-running cameras and phones from their own sound: how often a short window fixes the offset to within one 30 fps frame on real devices, rooms and clocks, and whether a confidence calibrated in one room or rig keeps its failure rate in another.

Step 1 (2026-11-13) runs on LibriWASN (Schmalenstroeer, Gburrek and Haeb-Umbach 2023; Zenodo, CC BY 4.0): five smartphones (Nexus, Pixel 6a, two Pixel 6, Pixel 7, Xiaomi) recording the same loudspeaker playback of LibriSpeech, 10 ten-minute sessions in a room with about 200 ms and 10 in one with about 800 ms of reverberation. Only the phone channels are fetched, by HTTP range requests, and none is committed. For each of the 200 phone pairs a reference line (20 s GCC windows every 10 s, Theil-Sen) is checked without ground truth: every pair kept at least 97% of its windows within 5 ms of the line, and the three lines of every phone triple add up to within 1.8 ms (median 0.3 to 0.5 ms), far under the 33 ms frame.

The phones' sampling clocks differ by up to 19.1 ppm (median 13.7 ppm), the same within about 1 ppm from session to session and room to room. Short windows (0.5 to 4 s, 40 seeded times per pair and session, a simulated start-offset prior of ±1 s because the published files are pre-aligned) were scored for five weightings, with the recordings as published and with independent white noise added at 0 and −10 dB. With added noise, PHAT is the worst weighting in both real rooms and partial whitening (gamma 0.25 to 0) the best: at −10 dB, 2 s windows fail 16.9% and 33.6% of the time with PHAT against 5.0% and 7.2% with gamma 0.25 (95% session-bootstrap intervals separate), which reverses capture-reverbsync's finding in simulated rooms.

Show the rest of the summary (32 more paragraphs, 2,396 words)

The 800 ms room is harder than the 200 ms one at −10 dB, as predicted. As recorded, PHAT still fails 3.3% of 2 s windows in the 200 ms room (spread over all ten pairs at scattered wrong lags, cause open; gamma 0.25 removes nearly all of it) against a predicted under 1%. The peak-to-sidelobe ratio separates good from bad windows (AUROC 0.91 to 0.998 at 2 s), but a threshold set in one room to let at most 1% of accepted windows be wrong, applied in the other, holds on average over noise levels (0.6% to 1.6%) and not at −10 dB from the drier to the wetter room (3.7% for gamma 0.25, 5.5% for PHAT).

Limits: one dataset, one loudspeaker playback of read speech, white rather than real noise, simulated start offsets, phones not cameras; the reference's own error (closure up to 1.8 ms) bounds sub-ms claims.

Step 2a (2026-11-13) tests the threshold theory of time-delay estimation (H4). Ianniello's (1982) gross-error model is extended to weighted GCC (weight (GxxGyy)−γ/2), fed with each window's Welch spectra, and checked on 84,000 simulated windows: Gaussian signals, and real read speech (MS-SNSD, MIT) with white or real babble noise in dry rooms and statistical rooms of RT60 0.2 and 0.8 s. zenodo.org was unreachable from that session, so the LibriWASN check is step 2b, with its harness committed.

On speech, the model predicts the share of windows more than a frame off within a factor of 2 in 137 of 171 scored cells (80%, against a pre-registered 75%). It runs low (median predicted/observed 0.58 to 0.74), misses most at 0 dB with 0.5 s windows, and agrees in 82 of 83 cells at −15 dB. The Gaussian check agreed in 32 of 36 cells, narrowly missing its pre-registered 90%.

Step 1's GCC normalises each FFT bin by its own magnitude, and that costs PHAT: averaged over the speech cells it misses 39.7% against 32.6% for the same weighting from smoothed spectra. With real babble the weighting ranking flips: PHAT is best and plain correlation worst (0.8 s room, −10 dB, 2 s: 50% against 79%).

Step 1's preference for partial whitening therefore holds for white noise only. As a blind confidence, the model's probability flags PHAT failures better than the peak-to-sidelobe ratio (median AUROC 0.85 against 0.74), but not for gamma 0.25.

Limits: the rooms are statistical, the delays are simulated, the babble is independent per device rather than shared, and the H4 score uses spectra aligned at the true lag.

Step 2b (2026-11-14) runs the same model window for window on the LibriWASN phones of step 1 (96,000 windows; the replay reproduced all 96 of step 1's miss rates exactly). H4 is refuted on real devices: the model is within a factor of 2 in 41 of 94 scored cells (44%, against the pre-registered 75%; Rice's variant 39/94). It now runs high (median predicted/observed 2.2, no cell underpredicted by more than x2), and its error grows with window length: 24 of 24 cells agree at 0.5 s, 1 of 24 at 2 s and 2 of 22 at 4 s, for every weighting.

The 800 ms room agrees more often than the 200 ms one (52% against 35%), and raw-bin PHAT again misses more than smooth-weight PHAT at −10 dB (all 8 window-room cells). Why real recordings improve with window length faster than the theory says is open.

Step 3 (2026-11-14) puts a distribution-free threshold on the peak-to-sidelobe ratio: Learn then Test (Angelopoulos et al. 2021), set for at most 1% wrong among accepted windows at delta 0.1, calibrated in one condition and tested in another on the same 96,000 windows. For 2 s windows with gamma 0.25, room to room gives 1.45% and 0.57% wrong among accepted (keeping 97 to 98%), and leaving each phone out gives 0.49% to 1.30%; but calibrating on the recordings as published and at 0 dB and testing at −10 dB gives 4.65% and 7.25%, because on cleaner audio the threshold falls to the floor.

A gate calibrated on a mix of noise levels controls the mix, not the noisiest part. Requiring the PHAT and gamma 0.25 estimates to agree (post hoc) cuts the noise-shift failure to 1.43% and 2.57% at 2 s, though not at 1 s. Within one room 13% of session splits exceed 1% (the i.i.d. bound is 10%), because windows cluster by session. Two rooms, five phones and added white noise only.

Step 3b (2026-11-14) tests two noise-aware gates on the same windows. A blind SNR estimate, the spread of a window’s 32 ms frame energies, separates the noise levels almost completely (2 s medians 21, 5.2 and 1.2 dB as recorded, at 0 dB and at −10 dB). Gating with one Learn-then-Test threshold per calibration quartile of that estimate, and refusing windows below the calibration’s range, keeps wrong-among-accepted under 1% in every shift: 0.27% room to room at −10 dB (keeping 74%, against 3.38% for the pooled gate), and 4% of within-room session splits above 1% (pooled 15%).

Under the unseen noise level it lets nothing wrong through, by refusing 93% and 99% of windows, most of them right. Scoring windows with the threshold theory instead of PSR is a slightly better flag for PHAT (AUROC 0.987 against 0.977) but does not carry the bound (2.25% and 5.37% under the noise shift). The recipe: cover the noise range in calibration, gate per blind-SNR bin, refuse below it. Added white noise only.

Step 3c (2026-11-15) asks whether that recipe survives real noise. With LibriWASN out of reach, it uses real read speech (MS-SNSD, MIT) in statistical rooms (RT60 0.2 and 0.8 s), 186,000 windows, each clean pair scored under white noise and nine kinds of real recorded noise (air conditioner, airport announcements, babble, copy machine, munching, neighbour, door, typing, vacuum cleaner) at +10, 0 and −10 dB, with each device's noise drawn from files the other never hears. It does not survive.

The blind SNR estimate reads real noise at −10 dB as 3.3 to 16.6 dB (white: 1.1 dB), because frame-energy spread measures how steady the noise is, not how loud. So the gate calibrated on white noise accepts 99–100% of 2 s windows and lets 2.6% to 81% wrong through, for every real type in both rooms. Calibrating on the other nine types, then testing on the held-out one, mostly refuses everything (the pooled gate keeps nothing), yet still passes 80% and 93% wrong windows when the held-out noise is typing.

Requiring PHAT and partial whitening to agree holds for steady noise (0.0–0.3% wrong for air conditioner, airport and vacuum cleaner) but not for impulsive noise (87–90% for typing). No fixed weighting wins: at −10 dB PHAT misses 2% of 2 s windows under air-conditioner noise against 12% for gamma 0.25, but 84% against 30% under copy-machine noise. Two reruns were needed before the result, both logged: the first let the two devices share noise samples, and the second missed duplicated noise files.

Step 3d adds blind features that name the noise (frame-energy kurtosis and spectral flatness) as Mondrian bins with conformal novelty refusal: on held-out noise types the gate keeps 43–48% of windows instead of 10–18%, but meets 2% at −10 dB for only 3–5 of 10 types (Neighbor 66–75% wrong, Typing 86–97%), because 92–100% of a new type's windows look like the library's. Calibrated on the site's own noise it holds the bound pooled over noise levels (9–10 of 10 types) but not at −10 dB alone.

Limits: simulated rooms and delays, and noise independent per device rather than one source heard by both.

Step 3e (2026-11-16) gave the gate the 4 s before each window: a minimum-statistics noise floor (a simplified Martin 2001 tracker) as the bin. Under site calibration at −10 dB it met the 2% bar for 4 of 10 noise types in each room, the same as the window-only estimate; for neighbour noise, door slams and typing the window’s own estimate rates −10 dB windows as less noisy than 0 dB ones (AUROC 0.36–0.47), and the context lifts that only to 0.48–0.66.

An oracle that knows the true level finds no threshold that certifies 1% for 7 of 10 types at −10 dB and refuses them, so the guarantee a single-microphone rig can state is the failure rate over its own mix of noise, with wrong windows at most 2.4% of the worst stretch here.

Step 3f (2026-11-17) lets both devices hear one noise source, placed with the talker at random around devices 3, 10 or 30 m apart, and scores errors against the clock rather than the talker. Any sound both devices hear arrives within D/c of the clock offset, so at 30 fps a lock on shared noise cannot cost a frame while the devices are under 11.4 m apart. Inside that bound shared noise is signal: at −10 dB, 2 s windows miss 0% under one shared babble source against 59–64% with independent babble, shared noise beats independent noise for 9 of 10 types in each room, and the gate keeps 60–63% of windows instead of 23–24%.

The exception is noise that repeats inside the search range: three looped door-slam recordings give equal peaks, 24–27% of windows miss, and a gate calibrated without shared noise passes 19–20% wrong. At 30 m the talker’s own path difference exceeds a frame in 69% of layouts, so about 70% of windows miss at every noise level and the gate refuses everything; a spread-out rig needs a clap equidistant from its cameras or a correction for known positions. Free-field geometry with statistical impulse responses; noise fully shared or fully independent, not mixed.

Step 3g replays those 242,000 windows bit for bit with the second-peak ratio (runner-up correlation more than a frame from the peak, over the peak) as the confidence: calibrated on independent noise and tested on shared noise at 3 m, the margin (1 − ratio) passes 1.1% and 0.0% wrong among accepted door windows where PSR passed 19% and 20%, keeps 69% and 68% of all windows against 63% and 60%, and cuts wrong-and-accepted windows from 1.4–1.5% to 0.0–0.1% of the total. It does not help impulsive noise heard by one device only (typing, munching, neighbours): the 2% bound holds for 4–5 of 10 types whichever score is used, because there the wrong peak has no close rival.

Step 3h replays step 3c’s 62,000 windows bit for bit with GCC on envelope-normalised audio (each device’s window divided by its own 32 ms moving RMS): at −10 dB with noise heard by one device only, windows more than a frame off fall from 78% to 0.1% under typing, 30% to 0.1% under munching, 21% to 0.2% under a door and 67–74% to 8–9% under neighbours, and the 1% gate on its own PSR keeps 84–91% of impulsive-noise windows (plain PSR kept 4–5%) while meeting 2% for 9 of 10 types (plain 4).

Steady noise gets worse (air conditioner 12–16% to 21–23%, vacuum 3% to 6–8%), so a per-window choice between the two estimators is the next step. A sub-window consistency gate fails (5–7 of 10 types), because periodic keystrokes align at the wrong lag throughout the window.

Step 3i fits a per-window switch between plain and normalised GCC on the window’s frame-energy kurtosis: read speech is itself bursty, so the fitted threshold sends 94–100% of windows to the normalised estimator and the switch does no better than always normalising (mean failure over ten noise types at −10 dB 11.5% and 12.5% against 11.3% and 12.5%; plain 33–35%); choosing by the higher PSR is worse (17.5–18.9%), and a per-window oracle would reach 7.8–8.9%, mostly on babble.

Step 3j (2026-11-18) checks steps 3c-3i's simulated room against the real phones: the same nine MS-SNSD noise types and white noise added independently per phone to step 1's 2 s LibriWASN windows (168,000 windows; step 1's windows replayed bit for bit, 6 of 6 cells equal). Across the ten noise types the real −10 dB failure shares rank like the simulated ones (Spearman 0.98 and 1.00 for gamma 0.25 GCC, 0.89 and 1.00 for the envelope-normalised estimator), and the sign of normalisation's effect agrees for 9 and 8 of 10 types.

Normalisation takes typing from 65.7% and 79.1% of windows a frame off to 2.8% and 1.3% and halves the failure summed over types (2.78 to 1.33, 4.01 to 2.11), at a steady-noise cost larger than simulated (air conditioner +7.4 and +10.5 points). The twin is optimistic about level: the normalised estimator fails 13% and 21% of real windows on average against 11% and 13% simulated, and white noise 5-10% against under 1%. A single 1% gate on the normalised estimator's PSR keeps 78% and 65% of real −10 dB windows and holds 2% for 7 and 6 of 10 types (misses are speech-like noise); on plain GCC it refuses nearly everything.

Limits: noise added rather than recorded in the room, one loudspeaker playback, phones not cameras.

Step 3k (2026-11-19) fits the twin to the phones on white noise only, choosing its direct-to-reverberant ratio and a white noise floor on real sessions 0-4, and tests it on the nine noise types it never saw (124,000 simulated windows, step 3h's replayed 62 of 62). No setting matches: in the twin reverberation is either harmless or breaks noise-free windows too (8–9% at DRR −6 dB in the 800 ms room against 0.1–0.4% real), so the fitted twin's mean gap over the nine types grows from 5.0 to 20.4 points in the 200 ms room and stays 7.4 in the 800 ms room, though it still ranks the types (Spearman 0.94–1.00).

A 1% gate set wholly on twin windows nonetheless holds 2% wrong among accepted on real test sessions for 7 and 6 of 10 types, keeping 77% and 68%, as well as the gate set on real recordings (7 and 6, 78% and 65%); the fitted, too-reverberant twin's gate holds for 10 of 10 types in the 200 ms room (0.20% wrong overall, keeping 67%), and no twin fixes speech-like noise in the 800 ms room. Post hoc, one real window in twenty has only 52–59% of its power in GCC's band against 89–90% simulated, a lead for a per-phone microphone model.

Limits: a coarse two-knob grid, a white rather than recorded floor, the same dataset as step 3j.

Research log

Ongoing: each run adds a batch and an entry, newest first. Source: LOG.md.

2026-11-19 · Step 3k outcome: no DRR or white noise floor makes the room twin match the phones, so fitting them does not predict the nine noise types (gap 5.0 to 20.4 points in the drier room, 7.4 unchanged in the wetter); but a gate set wholly in the twin holds on real phones as well as one set on them, and the pessimistic fitted twin's gate holds for 10 of 10 types

Show 27 earlier entries

2026-11-19 · Step 3k plan (calibrate the room twin to the real phones: does a twin fitted on white noise alone predict the nine noise types, and does a gate set wholly in it hold on real phones?), written before the run

  • Why now. Step 3j found the statistical-room twin chooses the estimator correctly but is optimistic about level (real white noise at -10 dB fails 5.0-9.7% against 0.1-1.0% simulated), so a gate's threshold could not be set in it. Two of the twin's inputs were never measured against a phone: the direct-to-reverberant ratio (fixed at 0 dB in step 2a) and the recorded noise floor (none: a simulated window's pauses are digital silence plus tail). This step sets both from real sessions 0-4 and tests the twin on what the fit never saw. If a twin calibrated on one cheap recording condition (white noise) predicts the rest, a rig's gate can be set in simulation; if not, it must be set on real recordings.
  • Method, fixed in advance by experiments/real-devices/twincal.py. The twin is step 3c's run_chunk at 2 s with rir(drr_db=DRR) and, optionally, independent white noise added to each device's reverberant stretch floor dB below its power before the window's power is taken (a second generator, so the main stream is untouched and drr 0 with no floor replays step 3h bit for bit; tests/test_twincal.py). Fit, per room, on step 3j's real windows (re-collected unchanged), sessions 0-4 only, 240 twin trials per grid point (DRR +6, 0, -6, -12, -18 dB; floor none, 40, 30, 25, 20, 15, 10 dB): (1) label-free, for each DRR the floor whose as-recorded median snr_hat is closest to the phones'; (2) of those five pairs, the one with the least squared error over six white-noise failure shares (plain and fn; as recorded, white 0 and -10 dB). The nine MS-SNSD types are not used. Predict: the uncalibrated twin (step 3h's replay) and the calibrated one, 1000 trials per room, all 31 conditions. Score: each twin's raw failure for the nine types at -10 dB against the real phones (all sessions, as step 3j); and T4 (Mondrian-snr_hat LTT, alpha 1%, delta 0.1, step 3's PSR grid) calibrated wholly on twin windows (none plus all ten types at 0 and -10 dB) and tested on real sessions 5-9, each type at -10 dB.
  • Seen before writing this. Steps 1-3j's results, including step 3j's real per-type failure shares, which the fit does not use but which I know; the three unit tests on synthetic speech. No twin window with a DRR other than 0 dB or with a floor has been generated or looked at; no real window of this run has been looked at.
  • Predictions (gamma 0.25, 2 s, 30 fps). (a) Replay: 62 of 62 uncalibrated cells (plain, fn) equal step 3h's. (b) The fit chooses a DRR below 0 dB and a finite floor in each room. (c) The calibrated twin's white-noise failure (in-sample) is within 2 points of the phones' for both estimators at -10 dB in each room. (d) Primary: over the nine types and both estimators at -10 dB, the calibrated twin's mean |sim - real| is at least a third smaller than the uncalibrated twin's (5.0 and 7.4 points) in each room. (e) Ranking kept: Spearman over the ten types at least 0.8 for both estimators and fn's sign (2-point dead band) right for at least 8 of 10, in each room. (f) A gate set in the uncalibrated twin, fn, meets 2% wrong among accepted on the real test sessions for at most 5 of 10 types in each room (the optimistic twin admits too much). (g) A gate set in the calibrated twin meets 2% for at least as many types as step 3j's gate set on real phones (fn 7 and 6) in each room, and its pooled wrong-among-accepted over all real test windows is at most 2%. (h) Exploratory, no direction stated: the as-recorded snr_hat spread (5th to 95th percentile) of each twin against the phones'.

2026-11-18 · Step 3j outcome: the simulated room ranks the noise types like the real phones (Spearman 0.89-1.00) and gets fn's effect right for 8-9 of 10 types, but is optimistic about level; fn halves the real failure (typing 66-79% to 1-3%) at a steady-noise cost larger than simulated

  • Ran as planned, no code changed after the run. 168,000 windows in about 18 min on 4 cores; scoring under a minute. The figure was drawn once.
  • (a) Replay: held. 6 of 6 step 1 cells equal in failure share and count (e.g. 800 ms room, white -10 dB, 0.0725 of 4000).
  • (b) fn halves plain for at least 3 of 4 impulsive types: held, 4 of 4 in each room. Typing 65.7% to 2.8% and 79.1% to 1.3%; Munching 27.5% to 2.9% and 38.5% to 2.4%; ShuttingDoor 21.6% to 2.9% and 32.3% to 2.1%; Neighbor 51.8% to 14.4% and 69.3% to 23.5%.
  • (c) The air-conditioner cost transfers: held, +7.4 and +10.5 points (simulated +8.6 and +7.0); VacuumCleaner +3.2 and +7.6.
  • (d) Summed over ten types fn fails less: held (2.78 to 1.33; 4.01 to 2.11).
  • (e) No difference when easy (0.5 points): failed, in fn's favour. At 0 dB fn is better by up to 2.3 and 3.5 points on impulsive types (Typing 3.7% to 1.4% and 0.3%); as recorded 0.9% to 0.8% and 0.4% to 0.1%. In simulation the two were within 0.2 points at 0 dB; the real recordings have bursts of their own at 0 dB that the simulation's did not.
  • (f) T4, fn at least as many types at 2% as plain: failed (7 against 8, 6 against 10), kept more: held (78% and 65% against 2% and 0%). Plain's count is a gate that refuses: the pool includes -10 dB typing, where plain is confidently wrong, and the few typing and munching windows it accepts in the drier room are 78% and 29% wrong. fn's misses are speech-like: AirportAnnouncements 3.1-4.3%, Babble 4.9-7.3%, Neighbor 5.1-5.7%, CopyMachine 2.5% (wetter room).
  • (g) Exploratory, the twin ranks types (Spearman at least 0.8) and gets the effect's sign for 8 of 10: held in both rooms (plain 0.98 and 1.00, fn 0.89 and 1.00; sign 9 and 8). The sign misses: AirportAnnouncements (simulated fn -5 and -3 points, real +1 and +2) and white in the wetter room (real +2.5).
  • (h) Exploratory, the gap. fn's mean -10 dB failure is 13.3% and 21.1% real against 10.9% and 12.7% simulated; plain's 27.8% and 40.1% against 32.5% and 35.4%. Real white noise fails 5.0-9.7% against 0.1-1.0%; fn's impulsive residual 1.3-2.9% against 0.0-0.3%.
  • What it changes. The statistical-room twin was good enough to choose the estimator: every qualitative conclusion of steps 3h-3i (normalise; it fixes impulsive noise; it costs on steady noise; speech-like noise is what is left) holds on real phones. It is not good enough to set a failure bound: real windows fail more under fn and far more under white noise, so a gate's threshold must come from real recordings. Next: calibrate the twin's room (noise floor, DRR, late tail) to the real white-noise failure and test whether it then predicts real levels, and whether a gate set on the calibrated twin holds on the phones.

2026-11-18 · Step 3j plan (does fn carry from the simulated rooms to LibriWASN's real phones? a twin check of steps 3c-3i), written before the run

  • Why now. Steps 3c-3i found everything on MS-SNSD speech through statistical rooms with a simulated delay, and step 3i made the envelope-normalised estimator fn the default on that evidence. A rig is real phones: real clocks (13.7 ppm median, step 1), real rooms, real microphones and their own recorded noise. The programme's theme is whether a simulated twin predicts the real system, so this step runs the same noise types on step 1's real recordings and puts each real cell beside its simulated twin. zenodo.org answers this session (HTTP 200), so the phones are reachable again.
  • Method, fixed in advance by experiments/real-devices/noisereal.py. LibriWASN 200_0S and 800_0S, 10 sessions each, 10 phone pairs, step 1's reference lines and its 40 seeded window times per pair. The loop consumes step 1's random stream exactly (window time, prior error, white-noise draws for every level and window length, step 1's skips), so the 2 s windows and their white noise are step 1's bit for bit, and step 1's three gamma 0.25 cells per room are a replay check. On the same as-recorded 2 s pair, the nine MS-SNSD types at 0 and -10 dB window SNR, independent per phone (each phone from its own half of the type's files, noisetype.noises), from a second generator seeded by (session seed, pair, trial), scaled as step 1 scales white noise. Per window: plain gamma 0.25 and fn (gamma 0.25 on each phone's audio divided by its 32 ms RMS envelope), error and PSR, and snr_hat. 21 conditions x 8,000 windows = 168,000. Failure: more than a 30 fps frame from step 1's reference. Gate T4: Mondrian-snr_hat LTT, alpha 1%, delta 0.1, step 3's PSR grid, calibrate on sessions 0-4 of a room (all conditions), test each type at -10 dB on sessions 5-9. Twin: each real -10 dB cell beside step 3h's simulated one (RT60 0.2 s for the 200 ms room, 0.8 s for the 800 ms room).
  • Seen before writing this. Steps 1 and 3c-3i's results; a 2-trial smoke run of one 200 ms session (80 s, 420 rows), of which 23 rows were printed: most windows within 4 ms of the line, a Babble and a Neighbor -10 dB window off by 2.2 s and 0.9 s for plain (fn 0.30 s and 3 ms). Nothing else; no failure share was computed.
  • Predictions (2 s, gamma 0.25, -10 dB unless stated; "at or under 2%" counts a gate that accepts nothing). (a) Replay: 6 of 6 step 1 cells (as recorded, white 0 and -10 dB, each room) equal in failure share and count. (b) fn at least halves plain's raw failure for at least 3 of the 4 impulsive types (Munching, Neighbor, ShuttingDoor, Typing) in each room. (c) The steady-noise cost transfers: fn fails more than 2 points above plain for AirConditioner in each room (simulation: +8.6 and +7.0). (d) Summed over the ten types, fn fails less than plain in each room. (e) No cost when easy: as recorded and at 0 dB, fn within 0.5 points of plain for every condition in each room. (f) T4: fn meets 2% for at least as many types as plain in each room and keeps more windows on average. (g) Exploratory, the twin: across the ten types the real failure shares rank like the simulated ones (Spearman at least 0.8) for both estimators in both rooms, and the sign of fn's effect (2-point dead band) agrees for at least 8 of 10 types in each room. (h) Exploratory, no direction stated: the mean gap between real and simulated failure at -10 dB, per estimator and room.

2026-11-18 · Step 3i outcome: a kurtosis switch between plain and normalised GCC does no better than always normalising; speech is itself bursty, so the fitted threshold sends 94-100% of windows to fn

  • Ran as planned, no code changed after the run. 62,000 windows collected in about 9 min on 4 cores; scoring under a minute. The figure was drawn once.
  • (a) Replay: held. 62 of 62 raw cells equal step 3c's and 62 of 62 fn cells equal step 3h's.
  • k\* came out low (0.61 and 0.57, log10 kurtosis), below the median of clean speech (1.02 and 0.93): the rule sends 96-98% of noise-free windows and 62-100% of every noisy type to fn. Median kurtosis at -10 dB is lowest for AirConditioner (0.65-0.67) and highest for Typing (1.54-1.56); white, Vacuum, Babble and no noise all sit at 0.83-1.02.
  • (b) Within 2 points of the better estimator for 8 of 10 types: failed in the 0.2 s room (7), held in the 0.8 s room (8). Misses: AirConditioner (17.8% against plain's 11.4%; 22.4% against 15.2%), VacuumCleaner (6.0 against 2.6; 6.8 against 4.0), Neighbor in the drier room (10.6 against 8.4).
  • (c) Halves plain's failure for 4 of 4 impulsive types: held in each room (Typing 77.2% to 0.2%, 81.4% to 0.0%).
  • (d) Steady noise within 2 points of plain: failed. AirConditioner +6.4 and +7.2 points, VacuumCleaner +3.4 and +2.8; the switch recovers about a quarter of fn's cost.
  • (e) No cost when easy (none, +10, 0 dB within 0.5 points of plain): held, every cell within 0.2 points.
  • (f) T4 at 8 of 10: failed (7 and 6, as fn); kept 74% against fn's 74% and 73%, held.
  • (g) T3 at least fn's count: held (9 and 9; kept 62% and 60%).
  • (h) Exploratory, psrmax fails more than kurt summed over types: held (17.5 against 11.5 and 18.9 against 12.5 points of mean failure x 10). Higher PSR picks plain on many impulsive windows (Typing 33% and 40% wrong), because plain's wrong click-train peak is sharp.
  • (i) Exploratory, kurt within 3 points of the oracle's mean: failed (11.5 against 7.8; 12.5 against 8.9).
  • Post hoc: the estimators fail on different windows, mostly on speech-like noise. Babble at -10 dB in the 0.2 s room: plain alone fails 268 windows fn gets right, fn alone 71 plain gets right, both fail 319 of 1000. The oracle's gain over fn comes from Babble, CopyMachine and AirportAnnouncements (6-8 points each), not from steady noise (1-2 points), so a better switch is worth most where neither kurtosis nor PSR tells the two apart. Also post hoc: psrmax is the better switch on steady noise (AirConditioner 12.4% and 17.0%, near plain) and under T4 in the wetter room (8 of 10 types, keeping 26%).
  • What it changes. For a rig the default becomes simply fn: it loses 7-9 points on air-conditioner-only windows at -10 dB and gains 20-80 points on everything impulsive or speech-like, and a kurtosis switch buys back only a quarter of the loss. A window's frame-energy kurtosis mostly measures the talker's own pauses, so it cannot say whether the noise is bursty. A switch would need a feature of the noise alone (the kurtosis of the residual, or of a band where speech is weak), or a learned combination of both correlations. Next: fn on step 3f's shared noise and on LibriWASN's real phones, which matter more than the last few points on steady noise.

2026-11-18 · Step 3i plan (choose plain or envelope-normalised GCC per window from the window's own frame-energy kurtosis), written before the run

  • Why now. Step 3h's normalised estimator (fn) fixed impulsive noise at one device (Typing 78-80% a frame off to 0.0-0.1%) and cut speech-like noise by a third to a half, but cost on steady noise (AirConditioner 12-16% to 21-23%, VacuumCleaner 3% to 6-8% at -10 dB), so neither estimator is the one to ship. A burst-free window has nothing for the normalisation to remove; step 3d's frame-energy kurtosis says, from the window alone, whether it has bursts.
  • Method, fixed in advance by experiments/gate/noiseselect.py. Step 3c's 62,000 2 s windows replayed exactly (noisetype.run_chunk unchanged, gcc wrapped; bit for bit on a 2-trial chunk, all 62 rows' first 11 columns equal). At gamma 0.25 the wrapper records fn's delay and PSR (as step 3h) and gate.frame_kurtosis of each device (b on the stretch aligned with a, as step 3d); kurt = the larger. Selectors: kurt uses fn when kurt > k*, else plain; k* is fitted per room on sessions 0-4, all 31 conditions pooled, as the calibration kurtosis percentile (0, 1, ..., 100) that minimises the selected estimator's failures (ties lowest; fit_kstar). psrmax uses whichever has the higher PSR. oracle the better per window (the ceiling, not deployable). Raw failure shares on test sessions 5-9. Gates as step 3h (Mondrian-snr_hat LTT, alpha 1%, delta 0.1, step 3's PSR grid, calibrate 0-4, test 5-9 at -10 dB), each selector gated on its chosen estimator's PSR and scored on its own error; T3 (none plus the same type) and T4 (none plus all ten types).
  • Seen before writing this. Steps 3c-3h's results; the 2-trial replay check and timing, and the first four rows' new columns (clean windows, fn on the same lag, kurt 1.0-1.2); three unit tests (tests/test_noiseselect.py; the click-train margin was lowered from 0.5 to 0.3 after the first run showed 0.80 against 0.44, a test of the feature on synthetic data, not MS-SNSD). No failure label was looked at with kurtosis in this run; step 3d used kurt as a gate bin, not as an estimator switch.
  • Predictions (gamma 0.25, 30 fps, test sessions; "at or under 2%" counts a gate that accepts nothing, coverage beside). (a) Replay: 62 of 62 raw cells equal step 3c's and 62 of 62 fn cells equal step 3h's. (b) At -10 dB, kurt's raw failure is within 2 points of min(plain, fn) for at least 8 of 10 types in each room. (c) kurt at least halves plain's raw failure for 4 of 4 impulsive types in each room. (d) Steady noise: kurt within 2 points of plain for AirConditioner and VacuumCleaner in each room. (e) No cost when it is easy: on no noise and at +10 and 0 dB, kurt within 0.5 points of plain for every type. (f) T4 (one generic calibration): kurt at or under 2% for at least 8 of 10 types in each room (step 3h: fn 7 and 6, plain 6 and 7), mean kept at least fn's minus 5 points. (g) T3: kurt meets 2% for at least as many types as fn (step 3h: 9 and 9). (h) Exploratory: summed over the ten -10 dB types, psrmax fails more than kurt in each room. (i) Exploratory: kurt's mean -10 dB failure within 3 points of oracle's in each room.

2026-11-18 · Step 3h outcome: envelope-normalised GCC fixes impulsive noise at one device (typing 78% to 0.1% a frame off; the 1% gate keeps 84-91% of those windows, not 4-5%) at a cost on steady noise; the sub-window gate fails

  • Ran as planned, no code changed after the run. 62,000 windows collected in about 11 min on 4 cores; scoring under a minute. The figure script was redrawn once after the first render (bars of wrong-among-accepted from 1-2 accepted windows hid the result); the numbers are unchanged.
  • (a) Replay: held. 62 of 62 raw failure cells (both weightings) equal step 3c's.
  • (b) fn halves raw failure for at least 3 of 4 impulsive types: held, 4 of 4 in each room. Typing 78.1% to 0.1% and 80.4% to 0.0%; Munching 29.6% to 0.1% and 29.3% to 0.3%; ShuttingDoor 21.3% to 0.2% and 23.1% to 0.1%; Neighbor 67.2% to 8.8% and 73.8% to 8.4%.
  • (c) fn costs little elsewhere: failed. No noise (0.0%) and white (0.2% to 0.1%, 0.7% to 1.0%) held; AirConditioner 11.9% to 20.5% and 15.7% to 22.7%, VacuumCleaner 2.5% to 5.5% and 3.4% to 7.8%. At +10 and 0 dB no type rises by more than 0.2 points. Speech-like and coloured noise also improve (Babble 59-64% to 39-44%, CopyMachine 30-36% to 14-18%).
  • (d) T3, sub2 at or under 2% for 8 of 10: failed (7 and 5). (e) sub2 keeps 10% of impulsive windows: failed in the 0.2 s room (9%), held in the 0.8 s room (12%).
  • (f) T3, fn beats psr: held. 9 against 4 types in each room; mean kept 60% against 24% and 23%; impulsive types kept 84% and 91% against 5% and 4%. Its misses: Babble 8.9% (9% kept) and AirConditioner 5.9% (7% kept).
  • (g) Exploratory, sub-window agreement ranks Typing failures at AUROC 0.8: failed (0.50, 0.53). 96% of failed typing windows have at least one sub-window on the same wrong lag.
  • (h) Exploratory, T4 sub2 at 8 of 10: failed (5 and 5). Under T4 fn meets 7 and 6 (Neighbor 6.7-7.5% wrong while keeping 95-96%): one generic calibration lets its confident speech-like errors through.
  • Post hoc: requiring plain and normalised GCC to agree, gated on fn's PSR, is worse than fn alone (8 and 7 of 10, 43-44% kept).
  • What it changes. The impulsive case that steps 3c-3g could only refuse is an estimator problem, not a gate problem: stop the bursts outvoting the window and the right lag wins, and PSR then works as a confidence again. Sub-window consistency was the wrong idea because keystrokes are periodic: the wrong alignment holds across the whole window. Next: choose plain or normalised per window (a rule from the window's own frame-energy kurtosis, step 3d's feature, fixed in advance), then that pair on step 3f's shared noise and on LibriWASN's real phones.

2026-11-18 · Step 3h plan (impulsive noise at one device: an envelope-normalised estimator, and a gate on whether the window's sub-windows find the same lag), written before the run

  • Why now. Step 3g closed the shared-noise aliasing case and left the impulsive independent one open: a key click, a bite or a door heard by one microphone gives a confident wrong peak with no close rival (PSR's AUROC for Typing 0.25, the second-peak margin 0.55), and step 3c found PHAT and gamma 0.25 land on the same wrong lag (the agree gate passed 87-90% wrong on Typing). Two ideas not yet tested: stop the bursts dominating the correlation before it is computed, and ask whether the lag holds across the window in time.
  • Method, fixed in advance by experiments/gate/noiseimpulse.py. Step 3c's 62,000 2 s windows replayed exactly: noisetype.run_chunk runs unchanged with its gcc wrapped (bit-for-bit on a 2-trial chunk: all 62 rows' first 11 columns equal). At gamma 0.25 the wrapper also records (1) fn: GCC on each device's audio divided by its own centred 32 ms moving-RMS envelope (sync.envelope_normalise, floor 1e-3 of the window's RMS), its delay and PSR; (2) three 1 s sub-window delays starting at 0, 0.5 and 1 s on the full window's lag axis (sync.subwindow_delays). Scores: psr (step 3c's), agree (step 3c's), sub2 (PSR if both halves land within a frame of the full estimate, else refused), sub3 (PSR if at least 2 of 3 sub-windows do), fn (the normalised estimator gated on its own PSR and scored on its own error). Gates as step 3g: Mondrian-snr_hat LTT, alpha 1%, delta 0.1, step 3's PSR grid, calibrate sessions 0-4, test 5-9, failure = more than a 30 fps frame from the speech truth. T3: calibrate on none plus the same type at all levels, test the type at -10 dB (step 3g's T3i). T4: one generic calibration on none plus all ten types.
  • Seen before writing this. Steps 3c-3g's results; the 2-trial timing and bit-for-bit check (and the first four rows' new columns, all clean-speech windows with sub-window offsets of 0); four unit tests, one of which shows raw GCC locking onto a repeated click template at each device and fn recovering the shared signal's lag. That test was rewritten once: its first version used independent random clicks, which do not correlate at all, so it tested nothing; repeated clicks (one key) are what can align. No failure label was looked at with the new features.
  • Predictions (gamma 0.25, -10 dB, 30 fps; "at or under 2%" counts a gate that accepts nothing, coverage reported beside). (a) Replay: 62 of 62 raw failure cells (both weightings) equal step 3c's. (b) fn at least halves the raw failure share for at least 3 of the 4 impulsive types (Munching, Neighbor, ShuttingDoor, Typing) in each room. (c) fn costs little elsewhere: no-noise failure within 0.5 points of gamma 0.25's, and white, AirConditioner and VacuumCleaner at -10 dB within 2 points, in each room. (d) T3: sub2 at or under 2% for at least 8 of 10 types in each room (psr: 4 and 4 in step 3g). (e) T3: sub2 keeps at least 10% of the impulsive types' -10 dB windows on average (not a gate that only refuses). (f) T3: fn on its own PSR meets 2% for more types than psr in each room and keeps more -10 dB windows on average. (g) Exploratory: the count of agreeing sub-windows ranks Typing failures with AUROC at least 0.8 in both rooms (PSR 0.25). (h) Exploratory: under T4, sub2 meets 2% for at least 8 of 10 types in each room.

2026-11-17 · Step 3g outcome: a second-peak ratio stops the looped door recordings PSR passed (19-20% wrong among accepted to 0-1%) and keeps more windows overall; it does not fix impulsive noise heard by one device only

  • Ran as planned. Collection about 20 min on 4 cores, scoring under 2 min. One fix after the run, before any number was read: the summary JSON failed to serialise numpy integers (default=lambda o: o.item()); the scoring code is otherwise unchanged.
  • (a) Replay: held. 720 of 720 raw failure cells and 240 of 240 PSR gate cells equal step 3f's.
  • (b) The ratio sees the aliasing: held. ShuttingDoor shared at 3 m, AUROC of the margin 0.91 and 0.91 against PSR's 0.73 and 0.71.
  • (c) T3s at 3 m, product 10 of 10 and more door windows kept: failed in the 0.2 s room (9 of 10; door 2.3% wrong among accepted), held in the 0.8 s room; the second part held (52% and 58% of door windows kept against PSR's 9% and 7%).
  • (d) T3 at 3 m, product 10 of 10, kept within 5 points: failed in the 0.2 s room (9; door 2.1%), held in the 0.8 s room; kept held (65% against 63%, 55% against 60%). The margin alone: door 1.1% and 0.0% wrong among accepted (keeping 56% and 51%) against PSR's 18.8% and 19.8%; 9 and 10 of 10 types (the miss is Typing at 2.02%, among 20% kept); 69% and 68% of all windows kept; wrong-and-accepted 0.10% and 0.00% of all windows against PSR's 1.50% and 1.42%.
  • (e) Impulsive independent noise, product at least two more types than PSR under T3i: failed (5 against 4 in each room; the margin alone 3 and 4).
  • (f) Exploratory, per above 0.8 for 30% of door windows: failed (4% and 3%); noise-free windows 0%, held.
  • (g) Exploratory, margin above PSR in AUROC for Munching, Neighbor and Typing under independent noise: held, 6 of 6, from a low base (Typing 0.55-0.56 against 0.25; Neighbor 0.62-0.63 against 0.59-0.60).
  • Post hoc: the product's grid was a poor choice. It reused PSR's grid (2 to 400); a product under 2 is refused outright, so it kept 14% of AirConditioner windows in the 0.8 s room where the margin alone kept 95%. The margin on its own grid is the better gate on these data; that is a post-hoc reading, to be confirmed on another set before it becomes the default.
  • What it changes. For shared sound the failure PSR missed was a close rival peak, and the second-peak ratio sees it at no cost in coverage. For impulsive noise at one device the wrong peak has no close rival (PSR even ranks those failures above successes, AUROC 0.25), so a different signal is needed: agreement between weightings or sub-windows. Next: that test, then a mix of shared and independent noise with the margin gate, and the margin on LibriWASN's real phones.

2026-11-17 · Step 3g plan (a second-peak ratio as the gate's confidence: does it catch aliasing on repeating sound and impulsive noise that PSR passes?), written before the run

  • Why now. Step 3f found shared room noise is signal inside a frame of sound travel except when it repeats: three looped door recordings put several near-equal peaks in the +-1.5 s search, and PSR, which compares the peak with the average sidelobe, passed them (T3 19-20% wrong among accepted at 3 m). Steps 3c-3e found the same kind of confident wrong peak under impulsive independent noise, which no level or class feature tamed. Both are a runner-up lag that nearly ties the winner; PSR cannot see a single strong rival. The second-peak ratio can.
  • Method, fixed in advance by experiments/gate/noiseperiod.py. Step 3f's 242,000 windows replayed exactly: noiseshared.run_chunk runs unchanged with its gcc swapped for sync.gcc_peaks, which returns gcc's delay and PSR bit for bit (checked on one 121-window trial, and by tests/test_noiseperiod.py) plus spr, the highest correlation value more than one 30 fps frame from the peak over the peak. Also recorded: per (gate.self_periodicity, the most a window repeats itself at lags of a frame to 1.5 s, the larger of the two devices'), a content-only flag. Three confidences through step 3f's gates (Mondrian-snr_hat LTT, alpha 1%, delta 0.1, calibrate sessions 0-4, test 5-9, labels against the clock at the D tested): psr (step 3f's), margin = 1 - spr (grid linspace(0.005, 0.995, 199)), prod = psr x margin (step 3's PSR grid). Calibrations T3 (independent noise of the type plus none) and T3s (shared noise of the type at that D plus none), tested on shared windows; T3i, the T3 gate tested on independent -10 dB windows.
  • Seen before writing this. Step 3f's results; the one-trial bit-for-bit check and its first five rows' per and spr values (speech-plus-noise windows at 0.06-0.16 and 0.14-0.81); the unit tests (a looped noise gives spr and per above 0.9, white noise per under 0.15). No failure label was looked at with the new features.
  • Predictions (gamma 0.25, -10 dB, 30 fps; "at or under 2%" counts a gate that accepts nothing, coverage reported beside). (a) Replay: every step 3f raw failure cell and all 240 of its psr gate cells (shared, -10 dB) equal. (b) The ratio sees the aliasing: for ShuttingDoor shared at 3 m, AUROC of margin for failure at least 0.90 in both rooms and above psr's. (c) T3s at 3 m: prod at or under 2% for 10 of 10 types in each room, and keeps more ShuttingDoor windows than psr (9% and 7%). (d) T3 at 3 m: prod at or under 2% for 10 of 10 types in each room (psr: 9 and 9), mean kept within 5 points of psr's. (e) Impulsive independent noise: T3i, prod meets 2% for at least two more types than psr in each room. (f) Exploratory: per above 0.8 for at least 30% of ShuttingDoor shared windows and at most 5% of noise-free windows. (g) Exploratory: margin's AUROC above psr's for Munching, Neighbor and Typing under independent noise in at least 4 of the 6 room-type cells.

2026-11-17 · Step 3f outcome: noise heard by both devices is signal, not noise, while the devices are closer than a frame of sound travel (11.4 m at 30 fps); beyond that the talker's own path difference already costs the frame, so geometry, not noise, sets the floor

  • Ran as planned, no code changed after the run. 242,000 windows collected in about 25 min on 4 cores; scoring under 2 min.
  • (a) Replay: held. 62 of 62 independent-noise failure cells (speech truth) equal step 3c's.
  • (b) Shared noise beats independent noise inside the frame bound: held. At D = 3 and 10 m, -10 dB, gamma 0.25, the shared-noise failure share against the clock is lower for 9 of 10 types in each room; the exception is ShuttingDoor (24% and 27% at 3 m against 21% and 23% independent). Under babble, independent noise misses 59% and 64% of windows; one shared babble source, 0.0%. The correlation lands within 2 ms of the noise source's own lag in 94-100% of shared -10 dB windows for 8 of 10 types (Munching 83-87%, ShuttingDoor 58-62%).
  • (c) At or under 2% for 9 of 10 types: failed, 8 of 10 in each room and at each of 3 and 10 m (ShuttingDoor 24-27%, Typing 2.4-3.1%).
  • (d) At 30 m shared noise fails more than independent for 5 of 10 types: failed, 3 of 10 in each room. Both fail about 70% of all windows, at every level including +10 dB, because the talker's own path difference exceeds a 30 fps frame in 69% of layouts at 30 m (the noise source's in 72%). Audio cannot sync to a frame there whatever the noise does; the noise only changes which wrong lag.
  • (e) The gate misses it at 30 m: failed, vacuously. Calibrated with clock labels at 30 m, about 70% of calibration windows are wrong at every PSR, so LTT certifies nothing and the gate refuses every window, in both T3 and T3s. That is the right answer, not a miss.
  • (f) T3 at 3 m, 10 of 10 types and more kept than independent: failed on the first part, held on the second. 9 of 10 in each room; ShuttingDoor passes 19% and 20% wrong among accepted while keeping 79% and 72%. Mean kept at -10 dB 63% and 60% under shared noise against 24% and 23% under independent noise.
  • (g) Exploratory, 60 fps at 10 m: held in the 0.8 s room (3 types), not in the 0.2 s room (2). As at 30 fps and 30 m, the floor is the talker's geometry: 52% of +10 dB windows are more than a 60 fps frame off at 10 m.
  • (h) Exploratory, T3s beats T3 at 30 m: failed (both refuse everything).
  • Post hoc: the door failures are aliasing on looped recordings. Three of MS-SNSD's eight ShuttingDoor test files repeat themselves (autocorrelation 0.98 at 0.534 s, 0.94 at 1.365 s and 0.94 at 1.52 s), and the wrong estimates cluster at those lags (+-0.4 to 0.55 s, +-1.06 s, +-1.35 to 1.55 s). With a shared periodic sound the correlation has several equal peaks inside the +-1.5 s search, the failure capture-aliaswin found on periodic content. Calibrating on the rig's own shared noise (T3s) brings it to 12.8% and 3.0% wrong at 3 m (keeping 9% and 7%) and under 2% at 10 m. Whether a real repeating machine or a door slammed at a steady pace does the same is not tested here; the loop is an artefact of the corpus.
  • What it changes. For frame-accurate audio sync the useful split is geometric. Inside c x frame (11.4 m at 30 fps, 5.7 m at 60 fps) any sound both devices hear is signal, so the hard case of steps 3c-3e (independent noise at each device: a phone in a pocket, wind on one mic, a camera beside a generator) is the one to gate, and shared room noise needs no gate except when it repeats inside the search range. Beyond it, audio alone cannot meet a frame, and the right move is a sound made on the devices' perpendicular bisector (a clapper equidistant from both) or a geometric correction from known positions. Next: a mix (shared plus independent noise), a periodic-content flag (second-peak ratio), and the overlap subsets.

2026-11-17 · Step 3f plan (noise heard by both devices: one source, a geometric path difference), written before the run

  • Why now. Steps 3c to 3e gave each device its own noise recording, so noise only ever hid the speech. In a room both devices hear one air conditioner or one door, and the correlation can lock onto that common sound at its own lag. Step 3c's limits named this case as untested; step 3e's outcome named it as next.
  • The geometry, exact before any run. For devices D apart, any source's arrival-time difference is at most D / c (triangle inequality). A camera rig needs the clock offset, and a peak on any shared sound, speech or noise, lands within D / c of it. So at 30 fps (33.3 ms, 11.4 m of travel) a lock on shared noise cannot cost a frame when D < 11.4 m; at 60 fps when D < 5.7 m. Beyond that it can, and PSR has no reason to flag it.
  • Method, fixed in advance by experiments/gate/noiseshared.py. Step 3c's 2 s windows replayed exactly (62,000 independent-noise windows, checked bit for bit on a 2-trial smoke run), plus shared noise on the same clean pairs: devices at (+-D/2, 0), talker and noise source uniform in [-D, D]^2 (one unit layout per trial scaled to D = 3, 10, 30 m), one noise stretch per type through a fresh statistical impulse response per device, device b's copy shifted by the noise's path difference relative to the talker's (whole samples), scaled to each device's speech power at +10, 0, -10 dB. 180,000 shared windows. All new draws from a generator seeded by (seed, trial, 3). Errors scored against the clock truth (err + the talker's path difference; sign checked by tests/test_noiseshared.py), 30 fps frame (60 fps secondary); independent windows scored the same way with the same trial's talker. Gates: step 3b's Mondrian-snr_hat LTT (alpha 1%, delta 0.1), calibrate sessions 0-4, test 5-9: T3 calibrated on independent noise of the same type (none + that type, all levels), T3s calibrated on shared noise of that type at the same D.
  • Seen before writing this. Steps 3c-3e's results; the 2-trial smoke run's timing and replay check (its failure numbers not printed); the two unit tests. The geometry alone, before the run: talker and noise path differences exceed a 30 fps frame only when D is above 11.4 m.
  • Predictions (2 s, gamma 0.25, -10 dB unless stated; "at or under 2%" counts a gate that accepts nothing). (a) Replay: 62 of 62 independent failure cells (speech truth) equal step 3c's. (b) Shared noise is signal inside the frame bound: at D = 3 and 10 m the raw failure share against the clock with shared noise is below the independent-noise share for at least 9 of 10 types in each room. (c) Inside the bound, shared noise rarely fails at all: at D = 3 and 10 m, raw failure at or under 2% for at least 9 of 10 types in each room. (d) Outside it: at D = 30 m, shared noise fails more often than independent noise for at least 5 of 10 types in each room (locks at the noise's own lag). (e) The gate does not see it: T3 at D = 30 m lets more than 2% wrong through among accepted for at least 5 of 10 types in each room. (f) Inside the bound, T3 at D = 3 m is at or under 2% for 10 of 10 types in each room and keeps more of the shared windows than of the independent ones (mean over types). (g) Exploratory, a direction only: at 60 fps and D = 10 m (bound 29 ms against a 16.7 ms frame) shared noise fails more often than independent for at least 3 types in each room. (h) Exploratory: T3s at D = 30 m meets 2% for more types than T3 in at least one room (calibrating with the rig's own geometry helps, mostly by refusing).

2026-11-16 · Step 3e outcome: no blind level estimate, even one that hears the 4 s before, carries the bound to -10 dB; for 7 of 10 noise types the true level says "refuse", and louder impulsive noise reads as cleaner audio

  • Ran as planned, no code changed after the run. 62,000 windows collected in 7.5 min on 4 cores; scoring 2.5 s.
  • (a) Replay: held. 62 of 62 raw failure cells and 31 of 31 snr_hat percentile cells equal step 3c's.
  • (b) ms orders levels: failed. AUROC of -10 against 0 dB windows is at least 0.90 for only 2 of 9 real types (AirConditioner 0.998, VacuumCleaner 0.990); it is above snr_hat's for 7 of 9 (the second half held). Speech-like noise reads 0.78 to 0.82 (AirportAnnouncements, Babble, CopyMachine) and impulsive or intermittent noise 0.48 to 0.66 (Munching, Neighbor, ShuttingDoor, Typing), close to chance.
  • (c) T3 site calibration at -10 dB, Mondrian-ms, 9 of 10 types and 25% kept: failed (4 of 10 in each room; 25% and 22% kept).
  • (d) Ceiling, the oracle at or under 2% for 10 of 10: held, but mostly by refusing. With the true level as the bin, LTT finds no PSR threshold that certifies 1% at -10 dB for 7 of 10 types in each room and accepts nothing; it keeps 55-100% for white, AirConditioner and VacuumCleaner, and 0% for the rest. So a level-conditional bound at -10 dB is reachable here only as a refusal for most real noise.
  • (e) ms meets 2% for more types than snr_hat under T3: failed (4 and 4 against 4 and 4).
  • (f) T1, knowing the level does not fix a change of kind: held. Calibrated on white noise, Mondrian-ms is over 2% for 8 of 9 real types in each room at -10 dB; so is the oracle (9 and 8 of 9).
  • (g) T2 leave one type out, Mondrian-ms under 8 of 10: held (3 and 7 of 10, keeping 12% and 9%).
  • (h) T0: held (0.10% and 0.10% wrong among accepted, keeping 99%).
  • (i) Exploratory, ctx between snr_hat and ms: held, barely (5 of 9 types).
  • Post hoc, not predicted: louder impulsive noise reads as cleaner audio. For Neighbor, ShuttingDoor and Typing the window's own estimate rates -10 dB windows higher than 0 dB ones (AUROC 0.36, 0.36, 0.47). An interpretation, not tested here: once intermittent noise dominates the speech, every blind level estimate measures the noise's own ratio of bursts to background, which does not change with gain, and bursts over a floor is what speech over a floor looks like. A level feature can only rank levels where the noise is steady (white, air conditioner, vacuum).
  • Post hoc: the failures are rare in absolute terms. In the failing T3 cells the gate accepts 1% to 14% of the -10 dB windows, so wrong-and-accepted windows are at most 2.4% (ms) and 2.0% (snr_hat) of all -10 dB windows of any type, against 0.2% and 0.8% for the oracle. The conditional rate is broken; the share of a session lost is small.
  • What it changes. A per-level guarantee needs to know the level, and under intermittent real noise a single-channel blind estimate cannot know it, with or without context; where the true level is known, the correct gate at -10 dB refuses most real noise anyway. So the honest claim for a rig is the marginal one (step 3d: the failure rate over the site's own mix of conditions), plus the absolute share of wrong windows accepted in the worst stretch. Next: noise heard by both devices (shared noise), where wrong peaks at the noise's own lag are the expected failure, and a confidence score other than PSR (a second-peak or a two-weighting test) for the impulsive types.

2026-11-16 · Step 3e plan (a level estimate from the seconds before the window: minimum-statistics noise tracking as the Mondrian bin), written before the run

  • Why now. Step 3d found that site calibration keeps the bound pooled over levels but not at -10 dB alone, because the window's own blind SNR estimate (frame-energy spread) does not separate levels under real noise, so each bin mixes them. It named the fix to try: a noise floor tracked over the seconds before the window (minimum statistics, Martin 2001), which a deployed rig has heard.
  • Method, fixed in advance by experiments/gate/noiselevel.py and gate.min_stats_snr_db. Step 3c's 2 s windows replayed exactly (62,000 windows), each now preceded by 4 s of context: the same session's preceding speech through the same impulse responses, and the same noise recording's preceding 4 s at the window's own gain (white context from a second generator seeded by (seed, trial), so step 3c's random stream is untouched). Features, the smaller of the two devices': ms (minimum of the recursively smoothed periodogram, alpha 0.7, over 1.5 s, summed over the 100-4000 Hz band; SNR of the window against it; no bias compensation, a simplification of Martin's tracker, disclosed), ctx (step 3b's frame-energy spread over the 6 s) and snr_hat (the window alone, as before). Gates, LTT at alpha 1%, delta 0.1: step 3b's Mondrian gate (4 quartile bins, refuse below the 1st percentile) on each feature, and an oracle with one threshold per true level (not deployable: the ceiling). Shifts T0 to T3 as step 3d.
  • Seen before writing this. Steps 3c and 3d's results; a two-trial check that the first 11 columns equal noisetype.run_chunk's bit for bit and that the context features are finite (1.5 s per trial); a 16-trial smoke run of collect and score with all output discarded. Unit tests on a synthetic bursty signal (monotone in white-noise SNR, compressed at the low end by the uncorrected bias); alpha was lowered from 0.85 to 0.7 there, before any MS-SNSD window was computed, because 0.85 did not reach the floor in a 200 ms pause.
  • Predictions (2 s, gamma 0.25 unless stated; "at or under 2%" counts a gate that accepts nothing as meeting it, coverage reported beside). (a) Replay: 62 of 62 raw failure cells and 31 of 31 snr_hat percentile cells equal step 3c's. (b) ms orders levels: AUROC of -10 against 0 dB windows at least 0.90 for at least 7 of the 9 real types, and above snr_hat's for at least 7 of 9. (c) T3 site calibration, -10 dB: Mondrian-ms at or under 2% for at least 9 of 10 types in each room, keeping at least 25% of the -10 dB windows averaged over types (step 3d's bar, missed by its class gate at 8 and 5 of 10, 13% and 14% kept). (d) Ceiling: the oracle meets (c)'s 2% bar for 10 of 10 types in each room. (e) T3, -10 dB: Mondrian-ms meets 2% for at least as many types as Mondrian-snr_hat in both rooms and more in at least one. (f) T1 white to real, -10 dB: knowing the level does not fix a change of noise kind: Mondrian-ms exceeds 2% for at least 3 of 9 types in each room. (g) T2 leave one type out, -10 dB: Mondrian-ms meets 2% for fewer than 8 of 10 types in at least one room. (h) T0: Mondrian-ms at or under 1% in both rooms (levels pooled). (i) Exploratory, a direction only: ctx's AUROC lies between snr_hat's and ms's for most types.

2026-11-15 · Step 3d outcome: features that name the noise widen what the gate keeps (10-18% to 43-48% of windows) but do not carry its bound to an unseen kind of noise; site calibration holds only pooled over levels

  • Ran as planned, no code changed after the run. 62,000 windows collected in about 4 min on 4 cores; scoring under a minute.
  • (a) Replay: held. 62 of 62 raw failure cells equal step 3c's.
  • (b) Kurtosis names impulsive noise: failed, 2 of 3. At -10 dB the median kurt (log10) is 1.55 for Typing and 1.42 for Munching against white noise's 95th percentile of 1.32; ShuttingDoor reads 1.30, just under. CopyMachine (1.36) is as "impulsive" as Munching. Flatness separates white noise (-0.2 dB) from every real type (-2.6 to -9.5 dB), but not the real types from each other.
  • (c) T2 class gate at or under 2% for 8 of 10 held-out types at -10 dB: failed (5 of 10 and 3 of 10, worse than mondrian's 7 and 6, which pass mostly by keeping nothing). Neighbor goes through at 66% and 75% wrong (keeping 44% and 37%), Typing at 86% and 97%, Babble at 21% and 13%.
  • (d) T2 coverage: held. The class gate keeps 43% and 48% of all held-out windows (types and levels pooled) against mondrian's 10% and 18%.
  • (e) Select: failed. It keeps a little more than class (48% against 43%, 51% against 48%) but meets the 2% bar for only 4 of 10 types in each room.
  • (f) T3 site calibration at -10 dB, class gate, 9 of 10 types and 25% kept: failed (8 of 10 and 5 of 10; 13% and 14% kept).
  • (g) T0 check: held. Class 0.06% and 0.15% wrong among accepted (keeping 87% and 68%); select abstains on everything in the 0.2 s room and is at 0.61% (16% kept) in the 0.8 s room.
  • (h) Exploratory, novelty refusal under T1 refuses more than half of the impulsive -10 dB windows: failed (1 of 3 in the 0.8 s room, Typing at 57%; 0 of 3 in the 0.2 s room, 25% to 41%).
  • Post hoc, not predicted: the novelty test almost never fires under T2. 92% to 100% of a held-out real type's -10 dB windows lie inside the other nine types' features; only white noise is refused outright (its flatness is unique). A library of real noise types spans the space of these three features, so "refuse what you have not seen" cannot see a new type.
  • Post hoc: site calibration does hold, pooled over levels. T3, all levels together (the distribution the bound is about): mondrian at or under 2% for 9 and 10 of 10 types, class for 9 and 10, select for 10 and 10. At -10 dB alone it fails because each bin mixes levels and the blind SNR estimate does not separate them under real noise: the guarantee is marginal over the calibration's mix of levels, not conditional on the hard ones.
  • Post hoc: select's design had a sample-size floor I missed. At delta/2 a bin needs 299 accepted calibration windows (230 at delta); the T0, T1 and T3 calibrations have about 2,000 windows over 8 bins, so most bins abstain. It is informative only under T2 (14,000 windows), where it picks PHAT in most bins.
  • What it changes. Blind single-window features (level, impulsiveness, flatness) neither identify an unseen kind of noise nor bin by level well enough under real noise, so a conformal bound for a rig is a bound on the site's own noise mix, and only marginally. Next: a level-conditional guarantee needs a blind SNR estimate that works under real noise (minimum-statistics noise tracking over the seconds before the window, Martin 2001, instead of one window's frame spread), and the shared-noise case (both devices hear one source) is still untested.

2026-11-15 · Step 3d plan (a gate that sees the kind of noise: kurtosis bins, conformal novelty refusal, per-bin weighting), written before the run

  • Why now. Step 3c showed the blind SNR estimate cannot place real noise and that impulsive noise gives confident wrong peaks, and named the fix to try: a feature that names the kind of noise, a weighting chosen per window, and calibration on the site's own noise.
  • Method, fixed in advance by experiments/gate/noiseclass.py and src/capture_sync/gate.py. Step 3c's 2 s windows replayed exactly (same MS-SNSD speech blocks, rooms, seeds and noise draws; 62,000 windows), with two blind features added per device: kurt (log10 kurtosis of 32 ms frame energies, the larger of the two) and flat (spectral flatness of the GCC band, dB, the mean). Gates, LTT at alpha 1%, delta 0.1: mondrian (step 3b, the baseline); class (8 bins, snr_hat quartile x kurt above/below its calibration median, plus conformal novelty refusal: 5-NN distance in scaled (snr_hat, kurt, flat), reference the even calibration trials, refusal above the 99th percentile of the odd ones; Bates et al. 2023); select (class's bins and refusal, the weighting per bin chosen on calibration between PHAT and gamma 0.25, each at delta/2). Shifts: T0 white to white, T1 white to each real type, T2 leave one type out, and T3 site calibration (calibrate on none plus the same type, other sessions).
  • Seen before writing this. Step 3c's results; a 16-trial smoke run to time it, check that the first 11 columns equal noisetype.run_chunk's bit for bit, and check the scorer runs (its output not read).
  • Predictions (to be reported held or failed; "at or under 2%" counts a gate that accepts nothing as meeting it, coverage reported beside). (a) Replay: all 62 raw failure cells equal step 3c's noisetype_summary.json. (b) The feature names impulsive noise: at -10 dB the median kurt of Typing, Munching and ShuttingDoor is each above the 95th percentile of white noise's. (c) T2, -10 dB, class with gamma 0.25: at or under 2% wrong among accepted for at least 8 of the 10 held-out types in each room (mondrian: 7 and 6). (d) T2: class (gamma 0.25) keeps a larger share of all held-out test windows (types and levels pooled) than mondrian in each room. (e) T2: select keeps at least as many windows as class (gamma 0.25) and meets (c)'s bar. (f) T3 site calibration, -10 dB: class (gamma 0.25) at or under 2% for at least 9 of 10 types in each room, keeping at least 25% of windows averaged over types. (g) T0: class (gamma 0.25) and select at or under 1% in both rooms. (h) Exploratory, a direction only: under T1, novelty refusal refuses more than half of the -10 dB windows of at least 2 of the 3 impulsive types in each room.

2026-11-15 · Step 3c outcome: a blind SNR estimate cannot tell real noise from clean audio, so step 3b's gate fails on every real noise type; a library of noise types mostly refuses, and still passes impulsive noise

  • Ran three times; the third run is the result. v1 drew both devices' noise from the same pool of recordings, so a pair could share noise samples and lock onto them at a wrong lag: among the 230 most confident 2 s windows of the pooled calibration, 41% were wrong. Found from that (post hoc), fixed in v2 by giving each device half of each type's audio; v2's confident-wrong share fell to 3.5%, but a check then found Neighbor_1 and Neighbor_7 are the same recording (normalised cross-correlation 1.0) and Munching_2/3/4/7 partly so (0.62 to 0.67). v3 gives each device whole files, duplicate groups kept on one side (split_files; noise_overlap_check.py finds no chunk of a's noise in b's above 0.50). The design otherwise did not change. v1 and v2 outputs are kept as noisetype_v1_results.txt and noisetype_v2_results.txt; every prediction below scores the same in all three runs, except that v1 met (d) only vacuously, by abstaining on every window.
  • (a) Real noise fools the blind SNR estimate: held, 9 of 9. At -10 dB, 2 s, white noise reads 1.1 dB (95th percentile 1.47); the real types read 3.3 (AirConditioner, VacuumCleaner) to 16.6 dB (ShuttingDoor), medians. Every real -10 dB window sits inside the white calibration's support.
  • (b) T1, step 3b's rule calibrated on white noise, at or under 2% fails for at least 3 of 9 real types in each room: held, 9 of 9 in both. 2 s, gamma 0.25, -10 dB: 2.6% (VacuumCleaner) to 77% (Typing) wrong among accepted in the 0.2 s room, 3.8% to 81% in the 0.8 s room, while keeping 99-100%.
  • (c) Pooled at least as bad: held (9 and 9 types over 2%, as mondrian).
  • (d) T2, leave one type out, mondrian at or under 2% for at least 7 of 10 held-out types in each room: failed (7 of 10 in the 0.2 s room, 6 of 10 in the 0.8 s room, and most of the passes accept nothing). Calibrated on the other nine types, the pooled PSR gate abstains on every window (its strictest admissible thresholds already let through more than 1%: 1.7% and 3.9% wrong among the 230 most confident windows, from Typing and Munching). Mondrian keeps a few bins: 10% and 18% of all T2 test windows. Where it accepts a held-out impulsive type it breaks the bound badly: Typing 80% and 93% wrong at 61% and 23% kept, Munching 50% and 38%.
  • (e) T0 in distribution: held (0.00%, 0.10%, 0.00%, 0.20% wrong among accepted at 2 s, keeping 75% (PHAT) and 99% (gamma 0.25)).
  • (f) Exploratory, agree beats mondrian under T1: held, 7 of 9 types in each room. It is near or under 1% for AirConditioner, AirportAnnouncements and VacuumCleaner (0.0% to 0.3%, keeping 70% to 91%) and fails on impulsive noise: Typing 87% and 90% wrong, Munching 38% and 33%. There the two weightings agree on the same wrong lag.
  • Post hoc, not predicted: no fixed weighting wins across real noise. 2 s, -10 dB, 0.2 s room, PHAT against gamma 0.25: AirConditioner 2.0% against 11.9%, AirportAnnouncements 5.9% against 25.2%, Babble 39.6% against 58.7%, Neighbor 27.6% against 67.2%, but CopyMachine 83.6% against 30.4%, Munching 68.0% against 29.6%, ShuttingDoor 34.2% against 21.3%, Typing 91.2% against 78.1%, VacuumCleaner 5.6% against 2.5%. Step 1's "partial whitening" holds for white and broadband machine or impulsive noise; for stationary coloured noise and speech-like noise PHAT is far better (step 2a's babble result, extended).
  • What it changes. Step 3b's recipe is a white-noise result. A frame-energy SNR estimate measures how steady the noise is, not how loud, so it cannot place real noise; and impulsive noise produces confident wrong peaks that no score built from one correlation flags. A gate that carries its bound needs a feature that names the kind of noise (impulsiveness, spectral flatness, or a learned noise class) and calibration on the site's own noise, or it must refuse. Next: a noise-class feature (spectral flatness and frame-energy kurtosis) as the Mondrian bin, a weighting chosen per window from it, and shared noise heard by both devices.

2026-11-15 · Step 3c plan (the noise-aware gate under unseen kinds of real noise), written before the run

  • Why now. Step 3b's deployable rule (calibrate with noise added down to the lowest SNR expected, one LTT threshold per bin of a blind SNR estimate, refuse below the calibration's support) was found with white noise only. A rig meets air conditioners, crowds, doors and typing, and step 2a showed real babble changes which weighting wins. This session's proxy refuses zenodo.org (403 on CONNECT), so LibriWASN is out of reach; github.com is not, so MS-SNSD (Reddy et al. 2019, MIT) supplies real speech and nine kinds of real recorded noise.
  • Method, fixed in advance by experiments/gate/noisetype.py. MS-SNSD clean_test speech split into 10 contiguous blocks ("sessions"), in step 2a's statistical rooms (RT60 0.2 and 0.8 s, a fresh impulse response per device), prior +-1 s, search +-1.5 s, failure more than a 30 fps frame off. 1000 trials per room and window length (1, 2, 4 s); each clean pair scored under 31 conditions: none, and white plus AirConditioner, AirportAnnouncements, Babble, CopyMachine, Munching, Neighbor, ShuttingDoor, Typing and VacuumCleaner at +10, 0 and -10 dB (independent stretches per device). Gates pooled, mondrian (step 3b) and agree (step 3), all LTT at alpha 1%, delta 0.1; calibrate on sessions 0-4, test on 5-9. Shifts: T0 white to white; T1 white (all levels) to each real type and level; T2 leave one noise type out (calibrate on none plus the other nine types).
  • Seen before writing this. Step 2a's and 3b's results; a 16-trial smoke run of the whole grid, timed and checked for columns and a clean scoring pass, with its numbers not printed. Nothing else.
  • Predictions (to be reported held or failed). (a) Non-stationary noise fools the blind SNR estimate: at 2 s, for at least 5 of the 9 real types the median snr_hat at -10 dB is above the 95th percentile of white noise at -10 dB. (b) T1, 2 s, gamma 0.25: the mondrian gate calibrated on white lets more than 2% wrong windows through at -10 dB for at least 3 of the 9 real types in each room. (c) T1: the pooled gate exceeds 2% at -10 dB (2 s, gamma 0.25) for at least as many types as mondrian, in each room. (d) T2, 2 s, gamma 0.25, -10 dB: mondrian stays at or under 2% for at least 7 of the 10 held-out types in each room (a library of real noise transfers to a new kind). (e) T0, 2 s, both gammas: mondrian at or under 1% in both rooms (the check). (f) Exploratory, a direction only: under T1 at -10 dB, 2 s, the agree gate lets through less than the mondrian gate for most real types.

2026-11-14 · Step 3b outcome: a gate conditioned on a blind SNR estimate keeps the bound everywhere, by refusing what calibration never saw; the theory score transfers no better than PSR

  • Ran as planned, no code changed after the run. 96,000 windows (gammas 1 and 0.25) collected in 17 min on 4 cores; scoring 9 s.
  • (a) Replay: held. 48 of 48 per-cell raw failure shares equal step 3's.
  • The blind SNR estimate separates noise levels cleanly. Median (5th to 95th pct) at 2 s: as recorded 21.1 dB (7.1 to 35.1), 0 dB 5.2 (2.7 to 6.4), -10 dB 1.2 (0.8 to 1.5). So in S4 only 8% (200 ms room) and 1% (800 ms room) of the -10 dB windows sit above the calibration's support floor.
  • (b) S4, 2 s, gamma 0.25, mondrian at or under 2%: held (0.00% in both rooms, kept 7% and 1%).
  • (c) It does so by refusing: held (keeps 7% and 1% of the -10 dB windows, of which 95% and 93% are right).
  • (d) Theory gate at or under 2% under S4 while keeping half: failed. 2.25% and 5.37% wrong among accepted, keeping 97%. Its threshold, like PSR's, is set where the calibration conditions are easy.
  • (e) S2 200 to 800, 2 s, gamma 0.25, -10 dB, both noise-aware gates at or under 2%: failed on theory (2.53%); mondrian 0.27% at 74% kept (pooled 3.38% at 95%).
  • (f) S1 cost, mondrian keeps at least 90% of pooled at 2 s, gamma 0.25: held (93% against 98%, 91% against 96%: ratios 0.95 and 0.95).
  • (g) Exploratory, theory beats PSR as a flag for PHAT at 2 s: held (AUROC 0.987 against 0.977, all noise pooled). For gamma 0.25 at 2 and 4 s PSR is the better flag (0.973 against 0.953; 0.987 against 0.968).
  • Post hoc, not predicted: mondrian also meets the within-room bound that step 3 missed. Over the 16 S1 cells (room x window x gamma), 4.0% of splits ended above 1% (pooled 15.3%, theory 21.5%), at a cost in kept windows (mean 65.8% against 72.2%). Per-bin calibration is more conservative, and the session clustering partly sits inside the noise bins.
  • What it changes. A failure bound under a noise shift needs calibration data at that noise level; no score fixes it alone. What the window can do is tell when it is outside the calibration, and the blind SNR estimate does that almost perfectly. So the deployable rule is: calibrate on recordings with noise added down to the lowest SNR expected, gate per blind-SNR bin, refuse below the floor. Next: overlap subsets (two talkers), and real babble, where frame-energy spread may not separate noise levels as cleanly as white noise.

2026-11-14 · Step 3b plan (a noise-aware gate: Mondrian bins on a blind SNR estimate, and a threshold-theory score), written before the run

  • Why now. Step 3 found that a Learn-then-Test gate calibrated on pooled conditions holds across rooms and phones but not when the noise level shifts (2 s, gamma 0.25, calibrated without -10 dB windows: 4.65% and 7.25% wrong among accepted against a 1% target). A deployed gate never sees the added-noise label, so the fix has to use something the window can measure. Two candidates: condition on a blind noise estimate (Mondrian calibration, with refusal outside the calibration's support), or replace PSR by a score that already models noise, the step 2a threshold theory evaluated blind.
  • Method, fixed in advance by experiments/gate/noise.py and src/capture_sync/gate.py. Step 1's windows replayed exactly (same sessions, seeds, times, prior errors, added noise) for gamma 1 and 0.25. Per window: error, PSR, p_blind (Ianniello cells model from Welch spectra of the two windows aligned at the window's own estimate, span 2.93 s) and snr_hat (10 log10 of the 90th over 10th percentile of 32 ms frame energies, the smaller of the two windows). Gates, all LTT at alpha 1%, delta 0.1: pooled (step 3), mondrian (one PSR threshold per calibration quartile of snr_hat; windows below the calibration's 1st percentile of snr_hat refused), theory (accept when p_blind <= t, t on a geometric grid of 241 values from 0.5 to 1e-6). Shifts S1 to S4 as step 3, S1 with step 3's splits.
  • Seen before writing this. Step 3's results; one 2-trial smoke run of one 200 ms session (240 windows, to time it and check the columns). Nothing else from this run.
  • Predictions (to be reported held or failed). (a) Replay: per-cell raw failure shares for gamma 1 and 0.25 equal step 3's summary.json replay to the digit. (b) S4, 2 s, gamma 0.25: mondrian keeps failure among accepted at or under 2% in both rooms. (c) It does so mostly by refusing: mondrian keeps under half of the -10 dB windows in both rooms (the -10 dB windows fall below the support floor). (d) S4, 2 s, gamma 0.25: the theory gate stays at or under 2% in both rooms while keeping at least half of the -10 dB windows. (e) S2 200 to 800, 2 s, gamma 0.25, -10 dB test: mondrian and theory each at or under 2% (pooled was 3.38%). (f) In distribution (S1, 2 s, gamma 0.25), mondrian keeps at least 90% as many windows as pooled in both rooms. (g) Exploratory, a direction only: on the real phones p_blind flags failures better than PSR (AUROC) for PHAT at 2 s, as on speech in step 2a (0.85 against 0.74).

2026-11-14 · Step 3 outcome: the gate's guarantee survives room and phone shifts on average but not a noise shift (2 s, gamma 0.25: 4.7% and 7.3% wrong among accepted against a 1% target)

  • Ran as planned, with one fix to the gate after a first scoring pass. 96,000 windows collected in 33 min on 4 cores. The first scoring pass abstained on every cell: walking from the strictest threshold, the first thresholds accept a handful of windows, where even zero failures cannot reject H0, so fixed-sequence testing stopped at once. Fixed by skipping thresholds that accept fewer than 230 windows (where zero failures first reject at alpha 1%, delta 0.1). The skip depends on scores only, not labels. The only outputs seen before the fix were "abstain everywhere" and the plug-in rule's numbers, which the fix does not touch. Tests added for both.
  • (a) Replay: held. All 120 step 1 failure shares reproduced to the digit.
  • (b) S1 within-room splits: held, with a caveat. Over windows and gates, 13.2% of splits had more than 1% wrong among accepted (bar 15%), against 33.1% for the plug-in rule. That is above the i.i.d. bound of 10% (gamma 0.25 alone: 19.8%), because windows cluster by session. No split went over 2%.
  • (c) S2 room shift, 2 s, gamma 0.25: held. Pooled over test noise, 1.45% (200 to 800) and 0.57% (800 to 200); at -10 dB, 3.38% (200 to 800), over 2%.
  • (d) S4 noise shift, 2 s, gamma 0.25: held. Calibrated without -10 dB windows: 4.65% (200 ms room) and 7.25% (800 ms room) wrong among accepted. The calibrated threshold sits at the grid floor, so the gate accepts every window, because on cleaner audio nearly all windows are right.
  • (e) S3 leave one phone out, 2 s, gamma 0.25: held, 5/5 at or under 2% (worst Xiaomi 1.30%, the only one over 1%).
  • (f) Agree keeps more than PSR under S2 at 2 s: failed. It keeps fewer (88% and 92% against 98% and 97%) and is wrong less often (0.70% and 0.53% against 1.45% and 0.57%).
  • Post hoc, not predicted: at 2 s the agree gate's threshold is the grid floor in every split, so it is effectively "accept when PHAT and gamma 0.25 agree". Under the noise shift it cuts wrong-among-accepted from 5.0% to 1.43% (200 ms room) and from 7.25% to 2.57% (800 ms room) at 84% and 68% kept. At 1 s it does not help: 4.7% and 11.3%.
  • What it changes. A gate calibrated on a pool of conditions controls the pool, not each condition. Rooms and phones are not where it breaks; the noise level, which a deployed gate does not observe, is. Next: calibrate per estimated window SNR (Mondrian, on a noise estimate the window can measure), and make the calibration session-level so S1 meets its own bound.

2026-11-14 · Step 3 plan (a Learn-then-Test gate and where its guarantee breaks), written before the run

  • Why now. Milestone 3. Step 1's PSR gate used a plug-in threshold with no guarantee, and it broke from the 200 ms to the 800 ms room at -10 dB (3.7%). A conformal threshold gives a finite-sample bound when calibration and test are exchangeable; this step measures where that bound holds on the real phones and where it fails.
  • Method, fixed in advance by experiments/gate/run.py and src/capture_sync/gate.py. Learn then Test (Angelopoulos et al. 2021): on a fixed geometric grid of 241 PSR thresholds (2 to 400), binomial p-values for H0 "failure among accepted > 1%", fixed-sequence from the strictest, delta 0.1. Data: step 1's windows replayed exactly (same seeds), with each window's error and PSR kept. Gates: PSR of PHAT, PSR of gamma 0.25, and "agree" (gamma 0.25's PSR when PHAT and gamma 0.25 land within a frame of each other, else refused). Shifts: S1 random 5/5 session splits within a room (50 per room); S2 room to room; S3 leave one phone out; S4 as recorded + 0 dB to -10 dB. Noise levels pooled in calibration except in S4. Step 1's plug-in rule on the same grid is reported beside it.
  • Seen before writing this. Step 1's aggregate results only. Nothing from this run; the gate code was tested on synthetic data only.
  • Predictions (to be reported held or failed). (a) Replay: the raw failure shares equal step 1's cells. (b) S1: averaged over windows and gates, the share of splits whose test failure among accepted exceeds 1% is at most 0.15 for LTT (bound 0.10 for i.i.d. windows; sessions are clusters), and higher for the plug-in rule. (c) S2, 2 s, gamma 0.25: pooled over test noise, failure among accepted stays at or under 2% in both directions; at -10 dB test noise it exceeds 2% in at least one direction. (d) S4, 2 s, gamma 0.25: calibrating without -10 dB windows and testing on them gives failure among accepted above 2% in both rooms. (e) S3, 2 s, gamma 0.25: at or under 2% for at least 4 of the 5 held-out phones. (f) Exploratory, a direction only: at 2 s the "agree" gate keeps more windows than PSR of gamma 0.25 alone under S2 in both directions.

2026-11-14 · Step 2b outcome: on the real phones the threshold model is refuted (41/94 cells within x2), right at 0.5 s and up to x5 pessimistic at 2-4 s

  • Ran as planned, no code changed. zenodo.org reachable; the phone channels of 200_0S and 800_0S fetched by range requests (one rename race from two fetches started by mistake into the same directory; the rerun skipped the complete files and fetched none). 96,000 windows, 4 gammas, 38 min on 4 cores.
  • (a) Replay: held. All 96 raw-bin miss rates equal step 1's.
  • (b) H4 on real devices: failed. 41 of 94 scored cells (44%) within x2 under the cells model, 39/94 under Rice's. H4 is refuted by the proposal's criterion (more than a quarter outside x2).
  • (c) Runs low: failed. It runs high: median predicted/observed 2.23, 53 cells over x2 and none under x0.5.
  • (d) 800 ms room agrees at least as often: held (24/46 against 17/48).
  • (e) Raw-bin PHAT worse than smooth-weight PHAT at -10 dB: held, 4/4 window lengths in both rooms.
  • Post hoc, not predicted: the error is a function of window length, for every gamma: 0.5 s 24/24 agree (median ratio 0.98), 1 s 14/24, 2 s 1/24, 4 s 2/22 (ratio 5.0). At 2-4 s PHAT agrees in 3/12 cells and the other weightings in 0/34.
  • What it changes. The model is usable as a miss-rate estimate for sub-second windows and as a pessimistic bound beyond. Step 2a's simulated rooms hid this, so the step 2a calibration does not transfer to real devices. Next: keep per-window outputs and find what the true-lag peak's L-scaling misses (speech activity, a segment-wise variance).

2026-11-14 · Step 2b plan (threshold theory on the real LibriWASN phones), written before the run

  • Why now. Step 2a tested H4 on real speech with simulated rooms and delays because zenodo.org was refused; this session's proxy reaches it (HTTP 200 on the record page), so H4 proper can run on the phones of step 1: real clocks, real rooms, real delays.
  • Design, fixed in advance by experiments/threshold/libriwasn.py (unchanged since step 2a merged). Replays step 1 exactly: 200_0S and 800_0S, 10 sessions each, 10 pairs, 40 seeded window times, prior error uniform in +-1 s, noise as recorded and independent white at 0 and -10 dB, windows 0.5/1/2/4 s. For gamma 1, 0.5, 0.25 and 0 it records step 1's raw-bin GCC, the smooth-weight GCC the model describes, and the cells and Rice predictions from Welch spectra of the two windows aligned at the reference line. A cell (room x noise x window x gamma) is scored if it has at least 10 observed failures; "agrees" means predicted/observed in [0.5, 2].
  • Seen before writing this. Step 1's and step 2a's results only. Nothing from this run.
  • Predictions (to be reported held or failed). (a) Replay check: the raw-bin GCC failure rates equal step 1's results.txt for the same cells (same seeds); if they do not, the run is not a replay and the rest is reported as such. (b) H4 on real devices, smooth-weight GCC, cells model: at least 75% of scored cells agree (the proposal refutes H4 if more than a quarter fall outside x2). (c) The prediction runs low, as on speech in step 2a: median predicted/observed below 1. (d) As in step 2a, the 800 ms room agrees at least as often as the 200 ms room. (e) Raw-bin PHAT fails more often than the smooth-weight PHAT in at least 3 of the 4 window lengths at -10 dB in each room.

2026-11-13 · Step 2a outcome: the threshold model predicts real-speech miss rates within x2 in 80% of cells, underpredicting; raw-bin PHAT adds misses; in babble PHAT wins

  • Ran as planned, with one change before any result was seen: the first full run was stopped after 22 min because np.convolve with 0.8 s impulse responses would have taken another hour. It was rerun with scipy.signal.fftconvolve, with the same seeds and the same maths. 84,000 windows, 4 gammas each, about 15 min on 4 cores.
  • (a) Gaussian implementation check: failed narrowly. 8 of 9 scored cells agree for every gamma (89%, the bar was 90%). The one miss is the same cell for all four gammas (-10 dB, 0.5 s), where the prediction is 0.15 to 0.17 against 0.07 to 0.08 observed. Median predicted/observed is 0.84.
  • (b) H4 on speech: held. 137 of 171 scored cells (80%) agree under the cells model, 143/171 under Rice's. The prediction runs low (median 0.58 to 0.74). The misses sit at high SNR and short windows (0 dB, 0.5 s: 22 of 22 underpredicted by more than x2), while 82 of 83 cells at -15 dB agree. The expected room pattern failed: dry agreed least (29/43), the 0.8 s room most (55/65).
  • (c) Raw PHAT beyond x2 of the prediction in most cells at -10 dB or below: failed (20 of 41 speech cells, 1 of 9 Gaussian). Raw-bin weighting does cost PHAT: its mean miss rate over the speech cells is 0.397 against 0.326 for smoothed PHAT, and 0.252 against 0.056 at -10 dB, 2 s, white noise, RT60 0.8 s. At gamma 0 raw and smooth are identical, as they should be.
  • (d) Exploratory: for PHAT, the theory probability from spectra aligned at the window's own estimate beats PSR as a failure flag in 12/12 scored 2 s speech cells (median AUROC 0.85 against 0.74). For gamma 0.25 it does not (0.567 against 0.578).
  • Post hoc, not predicted: with real babble (independent per device) the weighting ranking flips. PHAT is best and plain correlation worst (RT60 0.8 s, -10 dB, 2 s: raw PHAT 0.50, gamma 0.25 0.70, plain 0.79, smoothed PHAT 0.31). Step 1's "partial whitening, not PHAT" is therefore a white-noise result. Speech-shaped noise needs whitening.
  • What it changes. The gross-error rate is roughly predictable from a window's spectra once noise dominates, and the model gives PHAT a better confidence than PSR. The weighting default must depend on the noise spectrum (milestone 3's gate should see it). Step 2b, the LibriWASN run, is ready in experiments/threshold/libriwasn.py, pending a session that can reach zenodo.org.

2026-11-13 · Step 2a plan (threshold theory against Monte Carlo and real speech), written before the full run

  • Why now, and why not LibriWASN. Milestone 2 is H4: does Ianniello's (1982) gross-error model, fed with each window's measured spectra, predict how often a window lands on a wrong lag? This session's egress proxy refuses zenodo.org (403 on CONNECT; step 1's session could reach it), so the real-device recordings are out of reach. What is reachable: github.com, so MS-SNSD (Reddy et al. 2019, MIT): 37 min of real read speech (clean_test, 190 files) and real recorded babble (noise_test/Babble_*). This step therefore checks the theory where the answer is known (Gaussian signals) and on real speech and real babble with simulated delays and statistical rooms; the LibriWASN check (H4 proper) is step 2b.
  • Model (src/capture_sync/threshold.py). Correlator weight W = (Gxx Gyy)^(-gamma/2) from smoothed spectra; mean output at the true lag mu = max over +-1 ms of the integral of W Re(Gxy e^{j2 pi f tau}); spreads sigma1^2 = (1/2L) int W^2 Gxx Gyy and sigma0^2 = (1/2L) int W^2 (Gxx Gyy + Re Gxy^2) (Knapp and Carter 1976); largest wrong-lag output over 2.93 s of lag from M = 2 B_eff T independent cells (Ianniello) or from Rice's upcrossing rate. Spectra: Welch, 512-sample non-overlapping Hann segments, of the two windows aligned at the true lag. Estimators: the smooth-weight GCC the model describes (gcc_fixed) and step 1's GCC, which divides each bin by its own raw |X Y*|^gamma.
  • Grid. Gaussian white source, dry, white noise at -10/-15/-20 dB; real speech in dry, RT60 0.2 s and 0.8 s statistical rooms (a fresh impulse response per device and trial, DRR 0 dB), white or real babble noise at 0/-10/-15 dB. Windows 0.5/1/2/4 s; gamma 1, 0.5, 0.25, 0; 250 trials per cell; prior +-1 s, search +-1.5 s, failure more than 33.3 ms off, as step 1. A cell is scored only if it has at least 10 observed failures; "agrees" means predicted/observed in [0.5, 2].
  • Seen before writing this. One 60-trial probe per cell of Gaussian white, dry, gamma 0/0.5/1, L 0.5 to 2 s, -10 and -15 dB (the smooth-weight GCC failed 0.05 to 0.98; the theory said 0.16 to 0.81; raw PHAT failed 0.42 where the smooth one failed 0.08), and a 3-trial smoke run of the whole grid. Nothing else.
  • Predictions (to be reported held or failed). (a) Gaussian, smooth-weight GCC, cells model: at least 90% of scored cells agree, for every gamma (the implementation check). (b) H4 on speech, smooth-weight GCC, cells model: at least 75% of scored speech cells agree overall; I expect the dry cells to agree more often than the RT60 0.8 s cells, where reflections give the wrong-lag process structure the Gaussian model lacks. (c) Step 1's raw-bin GCC at gamma 1 (PHAT) fails more than twice the prediction in most scored cells at -10 dB and below, i.e. the raw-bin normalisation, not the smoothed weighting, is what makes PHAT worst; at gamma 0 raw and smooth are the same estimator and agree. (d) Exploratory, no prediction: whether the theory's probability computed blind (spectra aligned at the window's own estimate) flags failures better than the PSR of step 1 (AUROC).

2026-11-13 · Step 1 outcome: PHAT is the worst weighting in real rooms with noise; phone clocks differ by up to 19 ppm; a PSR gate transfers between rooms on average but not at -10 dB

  • Ran as planned, 20 sessions x 10 pairs x 40 times x 3 noise levels x 4 windows x 5 weightings = 480,000 estimates; no code changed after the run. All 200 references usable (at least 97% of 20 s windows within 5 ms of the line); triple closure median 0.5 ms (200 ms room) and 0.3 ms (800 ms room), max 1.8 ms.
  • Clocks. |SRO| median 13.7 ppm, max 19.1 ppm; each pair's rate repeats within about 1 ppm across sessions and rooms. The Nexus and Xiaomi sit 14 to 18 ppm off the three Pixels.
  • Predictions. (a) H1 at -10 dB, 2 s: held (PHAT 0.336 vs 0.169; gamma 0.25 0.072 vs 0.050, intervals touching). At 4 s the rooms are level. (b) H2: held in all four cells (e.g. -10 dB, 800 ms room: PHAT 0.336, gamma 0.25 0.072). (c) H3: held pooled over noise levels (1.6% and 0.6% among accepted), failed at -10 dB from the 200 ms to the 800 ms room (3.7%; PHAT 5.5%). (d) Failed: as recorded, PHAT fails 3.3% of 2 s windows in the 200 ms room.
  • Post-hoc, labelled as such. I expected (d)'s failures to come from one pair; bypair.py (4 s, 200 windows per pair) shows them spread over all ten pairs (PHAT 1.5% to 5.5%) at scattered lags, and gamma 0.25 removes nearly all of them. Cause open.
  • What it changes. For the programme: partial whitening, not PHAT, is the default from here; the gate needs a room-aware calibration (or a conformal bound with a stated exchangeability failure), which is milestone 3. For earlier projects: the ±40 ppm assumption of capture-syncgate, capture-syncslope and friends is about twice what these phones show.

2026-11-13 · Step 1 plan (real-device benchmark on LibriWASN phones), written before the full run

  • Why. Every earlier capture-*sync* result came from a simulated delay on one recording, with simulated drift and (but for capture-reverbsync) no room. LibriWASN has five phones recording the same playback on their own clocks in two real rooms (about 200 ms and 800 ms reverberation), so delay, drift and reverberation are all real.
  • Seen before writing this. One probe of session 0 of the 200 ms room with 3 window times per pair, recordings as published (no added noise): the ten reference lines had all or nearly all 20 s windows within 5 ms of the line, triple closure was 0.6 ms median and 1.3 ms max, clock-rate offsets were -16 to +15 ppm, and the published files start within 10 ms of each other. On that probe no 1 s to 4 s window failed. Nothing else.
  • Design. Reference per pair: 20 s windows every 10 s, PHAT, Theil-Sen line; usable if at least half the windows sit within 5 ms of the line. Short windows of 0.5, 1, 2 and 4 s at 40 seeded random times per pair and session; the other phone's stretch is cut around the reference delay plus a prior error uniform in +-1 s (the published files are pre-aligned, so the hand-start offset is simulated) and searched +-1.5 s. Noise: as recorded, and independent white noise added per phone at 0 dB and -10 dB window SNR (the reverberation and clocks are real; the noise is not). Weightings gamma 1 (PHAT), 0.75, 0.5, 0.25, 0. Failure: more than 33.3 ms from the reference. Intervals: bootstrap over the 20 sessions (10 per room).
  • Predictions (to be reported held or failed). (a) H1: at -10 dB, 2 s windows fail more often in the 800 ms room than in the 200 ms room, for PHAT and for gamma 0.25. (b) H2: at 0 dB and -10 dB, gamma 0.25 fails less than PHAT for 2 s windows in both rooms. (c) H3: a PSR threshold set in one room (accepted windows fail at most 1%, noise levels pooled) keeps accepted failures at most 2% in the other, for 2 s windows and gamma 0.25. (d) As recorded, 2 s windows fail under 1% in both rooms for every gamma.

Paper, code and results

More in Autonomous multi-angle capture

All 35 Autonomous multi-angle capture projects →

Other research topics

Proper scoring & elicitation28Effort & peer prediction9Verification games & slashing19Detection & tolerances15Markets as routers13Prediction-market design9Paying contributors8Decentralised training systems23Digital twins & sim-to-real164LLM chip design loops40Robot manufacturing loops53Ambient AI assistant46Self-replicating manufacturing34Simulation-first design64Engineered clothing45Task-specific AI accelerators40Algorithmic chip-design loops42Open-source purpose-built robots50The software-engineering loop43Adversarial human imitation41Open-source residential construction41Autonomous multi-angle capture35Sensing, imaging and 3D33Energy supply and dense energy storage41Robot societies among people32Agentic co-design of low-cost robots28Laundry-hanging robots26Machine learning on qualitative data23Robots on the building site28Civilizations & group psychology3History & human sciences4

Produced with AI assistance (Claude). Results are from the code in this repository and are simulated or stylised unless the abstract says otherwise; nothing here implies affiliation with or endorsement by any lab, researcher or company cited. MIT licensed.