Rustle: the repairs, and a second episode

Two things to check here. First, the nine rustles you tapped in story 6 — what was removed, and whether anything you wanted was removed with it. Second, whether the method finds anything on an episode it has never seen.

Five of the nine were replaced. Four were left alone because there is speech within 50 ms of the edge, and the rule is that a span containing any voice is refused rather than guessed at. Those four are below, untouched, for you to judge.

What replaced the noise is this episode's own quietest half-second of room tone, level-matched to the room either side of the gap. Not silence — digital silence in a pause is the exact fault we spent this whole build removing.

Every repair is length-preserving, and everything outside the five spans is bit-for-bit identical to before. Total audio touched: 1.45 seconds of 550.

nothing playing

The five replacements — before and after

Play BEFORE to hear the noise, then AFTER. The question that matters: is any speech missing or clipped in the AFTER version? Each starts 2 seconds early so you hear the run-in either way.

1 0:10.55 96 ms · was 13 dB above the room

2 1:52.09 107 ms · was 16 dB above the room

3 2:23.67 267 ms · was 12 dB above the room

4 7:44.33 331 ms · was 12 dB above the room

5 9:08.33 395 ms · was 16 dB above the room

The four left alone

These were refused automatically: there is voice inside or right at the edge of the span. Removing them risks taking a word or a breath with the noise. Your call whether they are worth a more careful attempt.

R1 0:24.05 320 ms · left alone — 10% speech in the span

R2 0:25.92 832 ms · left alone — 3% speech in the span

R3 2:21.76 714 ms · left alone — 3% speech in the span

R4 9:14.49 256 ms · left alone — 12% speech in the span

Second episode — story 5, never seen before

Nothing has been changed in this episode. These are the 25 highest-ranked candidates out of 431 seconds, found by separating the speech away and comparing what is left against real iPhone handling noise recorded by someone else entirely.

On story 6 the same method put 8 of your 9 known rustles in its top 60. This is the test of whether that was luck. Mark what you hear — if most are nothing, the method does not generalise and I would rather know now.

1 0:08.28 2.5 s

2 0:10.75 2.5 s

3 0:18.36 1.3 s

4 1:41.06 2.5 s

5 1:46.04 2.5 s

6 2:21.31 2.5 s

7 2:25.87 2.5 s

8 2:37.28 2.5 s

9 2:40.26 2.5 s

10 3:12.24 1.9 s

11 3:16.12 2.5 s

12 3:18.59 2.5 s

13 4:05.20 2.5 s

14 4:13.95 2.5 s

15 4:22.19 2.5 s

16 4:24.66 2.5 s

17 4:30.23 2.5 s

18 4:32.70 2.2 s

19 6:15.88 2.5 s

20 6:18.35 2.5 s

21 6:43.23 2.5 s

22 6:48.48 2.5 s

23 6:50.95 2.1 s

24 7:02.29 2.5 s

25 7:09.83 1.7 s

Your answers appear here — copy the block and send it to me.