Why Does My Instagram Transcript Generator Drop My Filler Words?

Because a clean transcript is a display decision, not a record of what reached the microphone. A filler, a stutter, a restarted half-sentence can be audible in the recording and still be absent from the text, for two different reasons that both get reported as “the tool cleaned it up”. This page separates those two reasons, reports what this site’s own word-level run on five public Reels actually measured, and gives a rule for keeping what was said.

Main guide: Instagram transcript — paste a link, get the full text. The page you are reading covers one specific case in depth.

Short answer: because disfluency removal is a separate display layer that runs after recognition and edits the result string. A Microsoft Azure speech writeup on display text format lists disfluency removal alongside inverse text normalization, capitalization, punctuation and content filtering as independent display layers, each rewriting the text passed to it, and notes that the inverse-text-normalization layer in particular is not user-configurable. A word can therefore go missing in two distinct places: the acoustic and language models may never emit it, or they may emit it and a later display layer may delete it for readability. Neither case means the speaker never said it. This site’s own word-level run on five public Reels (353.1 seconds, 1,114 words, faster-whisper base, per-word probability) found sentence-initial fillers present as decoded tokens — one quiet So at second zero — and explicitly set them aside from the confidence ranking; that is evidence the filler was in the recognized text before any editing decision was applied, not evidence it was absent from the audio. This page does not report a percentage of fillers removed, because that experiment had no independently listened reference transcript and did not run a layer-on / layer-off comparison. The practical fix is to choose verbatim or edited style before you quote anything, keep an untouched source copy, and replay the few seconds around any passage where a hesitation could change the meaning.

The silence in the text is a decision, not an absence

A recording is a continuous wave; a transcript is a string of tokens with spaces between them. Nothing in the wave is labelled “filler”. That label is applied later, by a component deciding what a readable line should contain. When a stutter, an um, a you know, or a half-started sentence disappears, the honest reading is not “the tool heard clean speech” but “some stage decided those tokens should not be displayed”.

That distinction matters because the two outputs answer different questions. A verbatim text answers “what sounds and words were in the recording”. An edited text answers “what would this look like as clean written prose”. Both are legitimate. Trouble starts when the second is quoted as if it were the first.

Four kinds of disfluent speech and what a display-layer edit does to each
What was in the audio What an edited text tends to show What the change actually is
Filled pause (um, uh, So at the start) Nothing, or a single discourse marker kept for flow A readability edit, not a recognition result
Repetition (I, I, I think) I think One or more decoded tokens deleted
False start (We need to — no, we should) We should A withdrawn clause removed from the result
Self-correction (on Tuesday, sorry, Wednesday) on Wednesday The earlier value dropped from the result

The last two rows are where a cleanup edit can quietly change facts: a withdrawn date, amount, or name is still something a speaker said, and a reviewer may need to see it.

What disfluency removal is

A Microsoft Azure speech writeup on display text format describes a stack of display layers applied after recognition: inverse text normalization (spoken “two thirty” written as “2:30”), capitalization, disfluency removal, punctuation, and content filtering. Each layer rewrites the text it receives. The writeup states that the inverse-text-normalization layer is applied and is not configurable by the user. A Chinese technical writeup on inverse text normalization as a sequence-labelling problem supplies the background for why these rewrites are treated as text-to-text decisions rather than as part of hearing the audio: the input is already a token sequence, and the layer decides what the written form of that sequence should be.

Disfluency removal sits in that stack as its own decision. It does not re-hear the recording; it acts on a sequence of words that someone or something has already produced. That is why deleting a stutter cannot be described as the model failing to hear it: by the time this layer runs, the hearing is done and the token is on the table. The layer is deciding whether the token belongs in written output.

This is the same family of layer that also adds punctuation and capitalization nobody said, and that replaces a flagged word with asterisks. Those pages cover the punctuation, casing and masking sides; this page covers the subtraction side — the layer whose whole job is to make a token disappear.

One layer adds, one layer removes. A reader sees a tidy sentence and assumes the tool was accurate. In fact the tidiness is produced by the removal layer; accuracy is a separate question that this page does not claim to measure.

Two different ways a word goes missing

Before blaming a display layer, separate the two stages, because they need different checks.

At the recognition layer. The acoustic model and the language model settle on the most likely word sequence given the sound. A quiet, short, unstressed filler at the very start of a clip is a weak acoustic target; a Chinese writeup on multi-accent data alignment and acoustic versus language-model error sources describes the two models disagreeing more under accented, fast, or mixed-language audio — the normal condition for a short Reel. So a filler can be dropped here, before any rewriting, simply because decoding did not settle on it. This is a word-selection outcome, not an edit.

At the display layer. The filler was decoded, then a disfluency-removal decision deleted it from the written result. This is an edit applied to a token that existed. It has nothing to do with whether the word was recognized correctly.

You cannot tell which happened by reading the final text — the output looks the same. You can only tell by looking at an earlier artifact (the raw recognition string, or a verbatim setting) or by replaying the audio. That is the practical reason to keep a source copy before cleaning anything.

What this site’s word-level run actually shows

The closest thing this site has to first-party evidence on fillers is a word-level decoding run on five public Reels: 353.1 seconds of audio, 1,114 words, 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, with nothing hand-corrected. The run’s own notes say sentence-initial fillers were excluded from the “first content word under 0.6” column, because a quiet So at second zero is low confidence for a boring reason.

That note is small but useful here. It means a filler was present as a decoded token in the recognizer output — present enough that the analysis had to make an explicit decision to set it aside. It was not missing from the recognized string. Whatever a later display layer would do to that So is a separate, downstream decision.

Word categories from the five-Reel word-level run and their low-confidence counts
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%

Function words — the class that holds discourse markers, hesitations and short connectors — were 305 of the 1,114 words, about a quarter of the text, with 19 low-confidence tokens. That is the population a cleanup layer operates on. Source: this site’s what breaks first run (0.6 is a fixed threshold chosen so the counts can be repeated, not a standard; confidence measures doubt, not error).

What was not measured

It is worth being exact about the limits, because a filler-removal rate would be easy to invent and would not be true.

  • The five-Reel run decoded straight into a recognition string. It did not run the same audio through a pipeline with disfluency removal enabled and disabled, so it cannot say how many fillers that layer would delete.
  • No run used an independently listened reference transcript. The text establishes what the recognizer returned, not what the speaker definitely said; that is why this page says the filler was in the recognized string, not that it was verified against a human transcript.
  • Nothing here measures Instagram’s own captions, tests a hosted speech service end to end, compares commercial products, or reports an accuracy rate.
  • No traffic, indexing or ranking figures appear on this page, because none were available to read.

So the claim on this page is deliberately narrow: a clean text is an edited artifact, the edit is a documented display layer, and the presence of a decoded filler token in this site’s own run is consistent with that. It is not a measurement of how often any product removes fillers.

How to keep what was said

You usually cannot disable the display layers, and you do not need to. You need to pick the artifact that matches the job before you quote anything.

  1. Choose the style first. Decide verbatim or edited before generating the text. A verbatim transcript keeps hesitations, restarts and repetitions; an edited one is prose. Quoting an edited line as a raw quote is where the damage happens.
  2. Keep an untouched source copy. Save the string exactly as the tool returned it, then make a separate edited copy. A reviewer should be able to see which words were removed.
  3. Treat a disappearance as an edit, not a fact. If an um or a restart is gone, do not conclude the speaker spoke fluently. Conclude that some stage decided not to show it.
  4. Replay the passages that carry meaning. A self-correction can withdraw a date, an amount, a name or a condition. Replay the few seconds around it and check the corrected value against the audio; the earlier, withdrawn value may still matter.
  5. Record the decision. Write which style you kept and what you removed, so a later reader knows the wording was chosen on purpose rather than by accident of which layer showed up.

Scope and limits

The layer model here comes from two sources: a Microsoft Azure speech writeup on display text format, which lists inverse text normalization, capitalization, disfluency removal, punctuation and content filtering as separate display layers and states the ITN layer is not user-configurable; and a Chinese technical writeup on inverse text normalization as a sequence-labelling problem, which frames these rewrites as text-to-text decisions. The word-level counts come only from one local experiment published on this site, described above with its parameters. Nothing on this page measures a product’s filler-removal rate, and no accuracy, traffic, indexing or ranking figure is reported.

FAQ

Why do my ums and stutters disappear from an Instagram transcript?
Two stages can remove them. The recognition layer may not emit a quiet filler, or a later disfluency-removal display layer may delete a token that was decoded. A Microsoft Azure writeup lists disfluency removal as one of several independent display layers applied after recognition. In neither case does the missing word mean the speaker never said it.
Does a clean transcript mean the speaker spoke cleanly?
No. A clean text is an edited artifact. Tidiness is produced by the removal layer; it is not a measurement of the audio. To know what was said, keep a verbatim or source copy and replay the passage.
Are fillers removed by the recognizer or by a display layer?
Either, and the final text looks the same. Only an earlier artifact or the audio can tell you which. This site’s five-Reel run found a sentence-initial filler present as a decoded token and set it aside from its confidence ranking, which shows the token existed in the recognized string.
Can I turn disfluency removal off?
Mostly no. The display layers are part of how the text is formed, and writeups describe the normalization layer specifically as not user-configurable. What you can control is the style you choose and the source copy you keep.
Did this site measure how many fillers were removed?
No. The five-Reel run had no independently listened reference transcript and did not compare the same audio with the layer on and off, so this page reports no removal rate. It reports only what the recognizer returned: 353.1 seconds, 1,114 words, 305 function words, 19 of them under the 0.6 threshold.
How do I keep the original wording?
Choose verbatim or edited style before generating, save an untouched source copy, treat each disappearance as an edit rather than a fact, replay passages where a correction changes a date or a name, and write down which text you kept and why.

How this connects to the rest of the site

Filler deletion is one member of the display-layer family; the neighbouring pages carry the other members and the measurements behind them. These three pages are the ones this page draws on most directly.

Affiliate link: Notta is a separate service that transcribes audio and video files you upload. This site earns a commission if you subscribe through that link, at no extra cost to you.