What breaks first in an Instagram transcript?

A name or a time expression. Names ran at 16.1% low confidence against 5.5% for the rest, and the first content word that wobbled landed in the first 18% of the clip on four of five Reels.

How can I find the weak spot without listening to the audio?

Ask the decoder for a probability per word and read the first one under 0.6, then scan for seams: a compound split in two, a time split into three.

Is 0.6 a standard threshold?

No. It is a fixed number chosen so the counts can be repeated. Report whatever threshold you use, because the totals move with it.

Why does the same name get different confidence each time?

Because the doubt is about where the word landed. One name scored 0.98 in the clear and 0.25 under a laugh, spelled the same both times.

Does a bigger model fix this?

Not tested here, so no claim is made. What was tested is where to look in the output you already have.

How do I stop a pipeline from breaking on repeated requests?

One request per link, cache the transcript, re-decode only the window you doubt. Re-fetching the same Reel for every check is what turns a working pipeline into a blocked one.

Instagram transcript: what breaks first, and how to keep it working

An Instagram transcript does not break at the words. It breaks at the first proper noun or the first time expression, and in this sample that happened inside the first quarter of the clip on four of five Reels. The repair is not a better tool. It is reading the first quarter first, and knowing the three shapes the failure takes.

“Breaks” here means one thing: the recogniser’s own per-word confidence drops below 0.6. The decoder exposes that number, so the weak spot can be found without listening to the audio and without a reference transcript. It measures doubt, not error.

What was measured

Five public Reels, decoded end to end: 353.1 seconds of audio, 1,114 words. Audio was extracted to 16 kHz mono and transcribed with faster-whisper base (int8, two CPU threads, English forced), asking for word-level timestamps and per-word probability. Nothing was hand-corrected. These are the same five Reels as the caption-layer test here, which read only the first ten seconds off the screen; this time the whole clip goes through the recogniser.

Reel	Length	Words	Under 0.6	First content word under 0.6
`Ddux1OTiIn7`	36.2 s	106	4	16.94 s — 47% in
`DdwIEyckv-Q`	95.8 s	324	23	2.00 s — 2% in
`Ddv4nWFjTiB`	62.9 s	208	18	6.80 s — 11% in
`Ddv7hsfNZnc`	90.2 s	221	11	15.82 s — 18% in
`DdwH6vtkfpm`	67.9 s	255	18	0.00 s — 0% in
Total	353.1 s	1,114	74	median 11% in

Sentence-initial fillers are excluded from the last column, because a quiet So at second zero is low confidence for a boring reason. What is left is the first word that carries content and still comes back shaky: four of five land in the first 18% of the clip. The exception is the shortest clip, comfortable for 16.9 seconds of 36 before wobbling on US.

Names fail about three times as often as ordinary words

Sorting all 1,114 words by what they are:

Kind of word	Count	Under 0.6	Rate
Capitalised names	93	15	16.1%
Tokens containing a digit	6	1	16.7%
Function words	305	19	6.2%
Everything else	710	39	5.5%

Names and numbers are 99 of 1,114 words, 8.9% of the text, holding 16 of the 74 doubtful tokens, 22% of the doubt. Doubt is not spread evenly; it pools on the tokens you cannot guess from context.

Two seams you can see without listening

The first is a compound split in two. In the 95.8-second news clip nor’easter came back as nor at 11.84 s (0.18) and Easter at 12.00 s (0.37): the first break in that clip, 12% in, visible as a space where no space belongs.

The second is time. In the 67.9-second interview clip all three time expressions split into three tokens each: five o ’clock, 2 .30, 7 .30. Every stray space sits where a human would have written one word or one figure, so a scan for spaces inside times and names catches both seams in seconds.

The same name, three confidences

One name appears three times in the 62.9-second podcast clip, spelled identically each time: Ed Davie. Its confidence was 0.58 at 6.80 s, 0.98 at 18.48 s and 0.25 at 59.82 s. Same string, same speaker, a four-fold spread. Confidence is not saying the name is hard; it is saying where the name landed — over a laugh, over a second voice. That is why the check has to be positional.

How to keep it working

Five steps, in this order. One: read the first quarter before you trust the rest; four of five first breaks sat there. Two: scan for seams — split names, split times, spaces inside words — which need no reference to spot. Three: check each name against the on-screen lower third, typed by a person, free to read. Four: when a name matters, re-decode the three to five seconds around it rather than the whole file, which is faster and leaves the rest stable. Five: one request per link and keep the result. Re-fetching the same Reel for every experiment is how a working pipeline starts returning nothing.

Scope and limits

Five Reels, English only, one recogniser on one laptop CPU, 353 seconds of audio. Confidence is not accuracy: a name the model has never seen can come back clean and misspelled. The 0.6 threshold is a choice made here so the counts can be repeated, not a standard; the totals move with it, the ordering between names and ordinary words does not. Nothing here says a larger model fails the same way, or measures Instagram’s own captions.

Questions people actually ask

What breaks first in an Instagram transcript?: A name or a time expression. Names ran at 16.1% low confidence against 5.5% for the rest, and the first content word that wobbled landed in the first 18% of the clip on four of five Reels.
How can I find the weak spot without listening to the audio?: Ask the decoder for a probability per word and read the first one under 0.6, then scan for seams: a compound split in two, a time split into three.
Is 0.6 a standard threshold?: No. It is a fixed number chosen so the counts can be repeated. Report whatever threshold you use, because the totals move with it.
Why does the same name get different confidence each time?: Because the doubt is about where the word landed. One name scored 0.98 in the clear and 0.25 under a laugh, spelled the same both times.
Does a bigger model fix this?: Not tested here, so no claim is made. What was tested is where to look in the output you already have.
How do I stop a pipeline from breaking on repeated requests?: One request per link, cache the transcript, re-decode only the window you doubt. Re-fetching the same Reel for every check is what turns a working pipeline into a blocked one.