Shmili episode 2 — the stutter, fixed

You heard words repeating. You were right, and the cause was bigger than that one spot. Here is the fix.

nothing playing

First — the stutter you heard. It was real, and it is fixed.

You were right. Words were repeating around 6:17.0. Here is that stretch before and after.

What went wrong: I had assumed the trimmed audio sat a fixed distance from the original tape all the way through. It does not. In two stretches of this episode it drifts — by about a sixth of a second from 6:13.1 to 6:41.6, and by a twentieth from 8:10.5 to 8:39.0. Ten of the twenty eight patches were therefore lifted from slightly the wrong place, and at a patch a fifth of a second long, being a sixth of a second out replays most of what you just said. That is the stutter.

Each patch is now taken from its own measured position. Both drifting stretches are about 28 seconds long, which points at the old cleaning tool processing in chunks and getting one chunk's position wrong.

Then — you asked whether the restored words were really in the original.

They were. Every patch is copied straight off the tape, sample for sample. Nothing was generated, invented, or re-voiced by any model. Here is the biggest one against what is on the feed today, turned up to match.

The whole episode.

This is the yes or no. Twelve and a half minutes.

Two small things.

The bedtime rule says no moment should jump more than 5 above the general level. With the music this sits at exactly 5.0 — passing, nothing to spare. Without music there is a little room:

Five of the eight sounds you marked still could not be removed. They sit on top of your voice and the only way to reach them is to cut into the voice. I left them. Say the word if you want me to try anyway. And 91 quieter sounds are still waiting if you want them.