Why Does My Instagram Transcript Generator Add Punctuation I Never Said?

Because a transcript is not a copy of the sound. After a recogniser has chosen the words, a second step rewrites them into written form: it inserts commas and full stops, decides which letters are capitals, turns spoken numbers into digits, removes hesitations and filters some words out. None of those marks are in the recording. On hosted speech services that rewrite is applied by the service and is not something you can switch off from the request.

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

Short answer: the layer has a name in speech documentation — inverse text normalization (ITN), also described as display text formatting. Recognition produces a word sequence; ITN converts that sequence from the way people speak into the way people write. Microsoft's Azure Speech documentation on display text format states that ITN is applied and is not user-configurable, and lists what it covers: dates, times, decimals, currency, addresses, email and phone numbers. That same documentation treats capitalisation, disfluency removal, punctuation and profanity filtering as separate display layers, each acting on the result text. So the comma in your transcript is a writing decision made by a text model after the audio was decoded, not a sound anyone made. The consequence for an Instagram Reel is practical: a fluent, well-punctuated transcript is not evidence that the wording is verbatim, and a “clean” sentence can hide a changed word. The fix is a separation pass — mark which characters could only have been added afterwards, replay the seconds around the words that matter, and keep the raw string. Below: what the layer changes, why it is not a setting, what this site's own measured runs actually returned, and a worksheet you can run in a minute.

What is in the recording, and what is added afterwards

Start by splitting the output into two columns. The waveform contains continuous sound: vowels, consonants, pauses, music, room noise. It contains no comma, no capital letter, no digit, no apostrophe, no paragraph break, and no full stop. Every one of those is a character in a text file, and characters have to be produced by something that writes text.

That something is a chain of post-recognition stages. A conventional pipeline decodes audio into a lattice of likely word sequences, picks one, and then applies formatting rules so the result reads like a document instead of like a transcript of speech. The distinction matters because people judge a transcript by how readable it looks — and readability is exactly the property the formatting layer was built to optimise.

Which transcript features exist in audio and which are added by text formatting
In the audio Only in the text Who decides
Vowels, consonants, stress, duration Commas, full stops, question marks A punctuation model using the surrounding words
A loud or quiet syllable Capital letters, sentence starts A capitalisation layer, not the loudness of the voice
A spoken number like “ninety-four” 94, 1994, 2.30 Inverse text normalization, choosing one written form
A short pause, a breath, a repetition No marker at all, or a silently removed “um” A disfluency-removal layer
A word the speaker did say An asterisk or a blank in its place A content-filter layer acting on the output

The right column is not heard. It is written. Azure's display-text-format documentation describes these as separate display layers applied to recognition results, with ITN not user-configurable.

Five written marks that nobody spoke

Once you know the layer exists, its output becomes visible. These five shapes show up in real transcripts, and each one has a different consequence for how you use the text.

  1. Punctuation. A full stop, a comma or an apostrophe is placed by a text model. In this site's eight-cut experiment, the same narration returned office in one run and office. in another, and asked against asked,, with the cut position as the only change. The mark moved; the audio did not.
  2. Capitalisation. A capital letter is a writing convention, not an acoustic event. The same experiment returned notebook and Notebook, and mangled a brand token into INST-Agram. and iNST agrom. depending on the cut. If you are quoting a product name, the casing in the transcript is not evidence of how the brand is spelled.
  3. Number format. Spoken numbers have no digits. Azure's documentation lists date, time, decimal, currency, address, email and phone as ITN categories. In the local measurements here, 94 became 1994 in one split run, and all three time expressions in one interview clip arrived as three tokens each: five o 'clock, 2 .30, 7 .30.
  4. Removed disfluencies. A hesitation, a repeated word or a false start can be deleted from the written result. The text then reads as if the speaker was more fluent than they were. That is a real loss if you are studying how someone said something, not just what they concluded.
  5. Filtered words. Content filtering acts on the result text, so a word can be replaced by a symbol or omitted. It is unrelated to whether the recogniser heard the word correctly, which is why a filtered transcript can look clean and still be incomplete.

None of these five is a recognition error. That is the awkward part: the layer is working as designed. A more accurate recogniser can still add a comma nobody said, and a cleaner-looking output can be further from the recording.

Why you cannot switch it off

“Can I turn off the punctuation?” sounds like a reasonable settings question. On a hosted speech service it usually is not one, because the formatting is applied by the service to the result it returns, not chosen by the caller at decode time. Azure's display text format documentation is explicit that inverse text normalization is not user-configurable; it also documents explicit punctuation and dictation modes as separate behaviours of the display layer. The practical reading is that you can choose a mode that changes how the layer behaves, but you cannot ask for a raw, unformatted word sequence from that endpoint.

There is a technical reason the layer sits where it does. Converting spoken forms to written forms is a sequence-labelling problem over text: given nineteen ninety-four, decide whether the output is 1994, 94 or two numbers, using context. That decision needs the surrounding words, so it belongs after recognition, not inside the acoustic model. A Chinese technical-community article on ITN as a sequence-labelling task describes exactly this framing. Once the step is downstream, it is a property of the service's output pipeline rather than a flag on your request.

This is also why two tools can return different-looking transcripts from the same Reel without either being wrong about the words. The word sequence may agree; the display layers differ. Comparing two outputs character by character and calling the differences “errors” will mislead you. Compare the words first, the marks second.

What this site's own measured runs returned

Everything below comes from two measurements already published on this site, both run locally with CPU speech models and no hand correction. They are not Instagram data and they do not establish how any hosted service behaves; they are here because they show the layer's effects on real output that can be re-checked.

Measured written-form changes from two local transcription experiments
Measurement What changed What it shows
Five public Reels, 353.1 s, 1,114 words Three time expressions split into three tokens each: five o 'clock, 2 .30, 7 .30 Stray spaces appear where a human would write one word or one figure
Same five Reels, word-level probability 6 tokens contained a digit; 1 of the 6 fell below the site's 0.6 threshold 1 ÷ 6 = 16.7%, on a sample far too small for a general rate
Eight cuts, 182.961406 s narration, 24 recognizer calls office → office.; asked → asked,; 94 → 1994 Punctuation and digits move with the cut, not with the audio
Same eight cuts notebook → Notebook; It's → Its; instructions → instruction Casing, apostrophes and word boundaries are written decisions
Same eight cuts, whole-file runs Each of eight runs returned 435 words and ended at 183.0 s against a 182.961406 s input A 0.039 s overshoot belongs to the whole-file axis only, not the split rows

Sources: the site's what breaks first measurement (353.1 seconds, 1,114 words, per-word probability, 0.6 threshold chosen here so counts can be repeated) and the eight audio cuts comparison (182.961406-second synthetic narration, 24 calls, cached faster-whisper tiny, CPU int8, English selected). The 16.7% figure is the arithmetic 1 divided by 6 on that specific sample; it is not an accuracy rate. The threshold is a local choice, not a standard.

One caution from those runs: a text difference is not automatically an error. Neither experiment used an independently listened reference transcript, so differences are recorded as differences. What they do establish is that the same recording can yield different written marks, which is the point of this page.

A worksheet that separates heard words from written marks

You do not need a better tool for this. You need one pass that labels the two columns before you quote anything.

  1. Keep the raw string. Before you tidy anything, save the transcript exactly as it came back. Formatting layers are lossy: once a hesitation is removed or a number is rewritten, the earlier form is gone.
  2. Highlight the characters that cannot be heard. Punctuation, capitals, digits, apostrophes, hyphens and line breaks. Do not treat them as evidence about the recording. Search for the components of a split figure as well as the whole string, because 2 .30 will not match 2.30.
  3. Replay around the words that carry meaning. Names, prices, dates, negations and anything you plan to quote. Three to five seconds is usually enough. Decide the word from the audio, not from how well the sentence reads.
  4. Check the frame for written forms. Captions, lower thirds and on-screen prices were typed by a person. For a figure or a brand, that printed form is stronger evidence than the recogniser's spelling.
  5. Record the decision, not just the edit. Write the confirmed form beside the original and keep the note. If the file goes to someone else, the note is the only proof that a mark was checked rather than assumed.

One rule above all five: if a formatting layer changed something you cannot verify, leave it visible. Silently normalising the text makes the output easier to read and harder to audit.

What a browser-based generator changes here, and what it does not

This site runs the recogniser in your browser and returns the decoded text as one string. There is no hosted display pipeline in that path, so the output does not arrive with a service-side punctuation layer applied to it. That is a real difference, and it is verifiable: the copy button reads from the same element the download button writes to a .txt file.

It is also a narrower difference than it sounds. A raw word sequence is not a verbatim record. The model still chooses between words that sound alike, still guesses how to spell a name, and still has no way to hear an apostrophe. A local generator removes one layer of writing decisions; it does not turn the remaining text into evidence. And if you need publishable punctuation, you will add it yourself — which is fine, as long as you know you are the one adding it.

What you can check

Was the word spoken at all?
Which of two near-sounding words?
Which kind of number: year / price / time?
What does the on-screen text say?
Did I keep the raw string?

What stays a writing decision

Where a comma or full stop belongs
Which letters are capitalised
Whether 94 was 1994
Whether It's or Its was meant
Where a hesitation was removed

The left column lists checks a person can finish against the audio or the frame. The right column lists marks that no recording contains.

Scope and limits

The mechanism described here comes from Microsoft Azure Speech documentation on display text format (ITN is not user-configurable; dates, times, decimals, currency, addresses, email and phone are covered; capitalisation, disfluency removal, punctuation and content filtering are separate display layers) and from a Chinese technical-community article framing ITN as a sequence-labelling task. The measured examples come only from two local experiments published on this site, described above with their parameters. 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.

FAQ

Why does my transcript have commas and capitals the speaker never said?
Recognition produces a word sequence, then a formatting layer rewrites it into written form. Punctuation, capitalisation, digit forms, removed hesitations and filtered words are all added or changed by that layer. They are characters in a text file, not sounds in the recording.
Can I turn off inverse text normalization on a hosted speech service?
Azure's display text format documentation states that ITN is not user-configurable. Some services offer modes such as explicit punctuation or dictation, but those change how the display layer behaves; they do not return an unformatted word sequence. The formatting sits downstream of recognition, where it has the surrounding words it needs.
Is a well-punctuated transcript more accurate?
Not necessarily. Readability is what the formatting layer optimises, and it can be applied to a word sequence that still contains the wrong word. A clean-looking sentence is not evidence that the wording is verbatim.
Why did the same Reel come back with different punctuation from two tools?
The word sequences may agree while the display layers differ. Compare the words first and the marks second; calling every character difference an error will mislead you.
Does a browser-based generator avoid this problem?
It removes the hosted display pipeline from the path, so the text arrives without a service-side punctuation layer. It does not remove the model's own choices between near-sounding words, spellings of names, or number forms. You still need the verification pass.
What should I do before quoting a transcript?
Keep the raw string, mark the characters that cannot be heard, replay the seconds around names and figures, check any on-screen text, and record the confirmed form beside the original. If a mark cannot be verified, leave it visible rather than silently normalising it.

How this connects to the rest of the site

Written marks are one layer among several. These three pages cover the measurements and the neighbouring cases this page draws on.

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.

Next case: why a transcript capitalizes the wrong words — the capitalisation layer guessing proper nouns from context.

Next case: why a transcript picks the wrong homophone — when the model swaps in a different legal word rather than a misspelling.

For the layer that removes words rather than adding them, see why a transcript drops filler words and restarts.