Why Does My Instagram Transcript Generator Capitalize the Wrong Words?
Because a capital letter is not in the recording. After a recogniser has picked the words, a separate step decides which ones start with a capital. It does not hear a louder voice or a “this is a name” signal — it guesses from the words around it. So when a brand, person or place name comes back with the wrong casing, or with a capital the speaker never stressed, that is the model writing, not the speaker. The quiet danger is that a wrong capital looks right: a name spelled “Iphone” or a brand returned as “INST-Agram” reads like a deliberate choice, so it survives proofreading that would catch an obviously wrong word.
Main guide: Instagram transcript — paste a link, get the full text. The page you are reading covers one specific case in depth.
Short answer: capitalisation is a display layer, not a sound. Speech documentation describes capitalisation as one of several display layers applied to recognition output — separate from the acoustic model — and on hosted services it is not a setting you control (Microsoft Azure's display-text-format documentation lists capitalisation among display layers and states inverse text normalization is not user-configurable). The capitaliser works from the surrounding words: it infers that a token is a proper noun and capitalises it. When the inference is wrong, you get a capital the audio never implied — a brand spelled “INST-Agram.”, a common word turned into “Notebook”, a contraction resolved as “Its” instead of “It's”. This site's own repeated transcription runs show the casing flipping with nothing in the audio changing. The fix is to treat every capital on a name as a guess, replay the seconds around it, and check the on-screen branding. Below: why the model guesses, what the measured runs actually returned, why a wrong capital is the most dangerous kind of error, and a worksheet you can run in a minute.
A capital letter is a writing decision, not a sound
The waveform carries pitch, loudness, timing, vowels and consonants. It carries no uppercase. A capital “A” is a character in a text file, produced by something that writes text. So capitalisation must come from a step that reads the chosen word sequence and decides casing — a text model, not the microphone. Azure's display text format documentation treats capitalisation as a display layer applied to recognition results, alongside punctuation, disfluency removal, number formatting and content filtering; it states that inverse text normalization is not user-configurable. In other words, the casing you see was written after the audio was decoded, by a layer you generally cannot switch off.
| In the audio | Added by the capitalisation layer | Consequence |
|---|---|---|
| A louder or stressed syllable | A capital at a word start | Loudness is not a flag for “proper noun” |
| A spoken name like “instagram” | Instagram, INST-Agram., iNST agrom. |
The model picks one casing from context |
| A sentence boundary (a pause) | A capital after a full stop | The pause is real; the capital is a writing rule |
| The word “it's” as heard | It's or Its, with or without a capital |
Apostrophe and case are spelling guesses |
The right column is written, not heard. Azure's display-text-format documentation describes capitalisation as a display layer applied to recognition results.
Why the model guesses the capital
A recogniser has no dictionary entry marked “capitalise me.” It returns a word sequence, and the capitaliser decides proper-noun status from context — the words before and after. That context-based guess is exactly where it can go wrong: a brand the model does not recognise gets a generic capital on the first letter, or gets split and each piece capitalised. A Chinese technical-community writeup on recognition under Cantonese / multi-model fusion and mixed-language practice notes that named entities are among the hardest cases for recognition — and capitalisation rides on top of that recognition. When the recogniser is least sure what the name is, the capitaliser is guessing the casing of a word it is not even sure about.
None of this is a “bug.” The capitaliser is doing its job: turning speech into readable text. The problem is only that its casing looks authoritative, so a reader trusts it as if it came from the recording.
What this site's own measured runs returned
Two local experiments published on this site, run with CPU speech models and no hand correction, show casing moving with no change in the audio. They are not Instagram data and do not measure any hosted service; they are here because they show the effect on real, re-checkable output.
| Measurement | What changed | What it shows |
|---|---|---|
| Eight cuts, 182.961406 s narration, 24 recognizer calls | notebook → Notebook |
Casing is a model choice, not in the audio |
| Same eight cuts | A brand token returned as INST-Agram. and as iNST agrom. |
One brand, two invented casings |
| Same eight cuts, contraction | It's → Its |
Apostrophe and case are spelling guesses the audio cannot settle |
| Five public Reels, 353.1 s, 1,114 words (site measurement) | Names were the spot-check failure point | Names are where capitalisation matters most and is least reliable |
Sources: the site's eight audio cuts comparison (182.961406-second synthetic narration, 24 calls, cached faster-whisper tiny, CPU int8, English selected) and what breaks first (353.1 seconds, 1,114 words). The casing differences are recorded as differences; neither run used an independently listened reference, so they establish that casing moves, not that any specific word was “wrong.”
The dangerous part: a wrong capital looks right
A wrong word usually looks wrong. “nor Easter” instead of “nor'easter” is different enough to notice — except the capitalisation makes both halves look like deliberate proper nouns, so a proofreader reads “Nor Easter” as a name and moves on. The site's main transcript guide records exactly this: a 1:36 clip returned “nor Easter” for “nor'easter”, the apostrophe gone and both halves capitalised as though it were a person. A mangled brand casing like “INST-Agram.” reads like a stylised logo, not an error. This is why capitalisation errors are the ones most likely to reach a published caption: they survive the edit that would catch an obviously wrong word. The rule is simple — any capital on a name is a claim about spelling that the audio did not make, and it deserves the same replay-and-check as a suspicious word.
Mixed language makes the guess shakier
Capitalisation rules are language-specific. A name that is lowercase in one language may be capitalised in another, and a mixed-language Reel gives the capitaliser two rule sets to choose between with no signal about which applies mid-sentence. The site's mixed-language coverage looks at this from the recognition side; the capitalisation layer inherits the same ambiguity. If the Reel switches languages around a name, treat the casing as doubly uncertain.
A worksheet: how to catch a wrong capital
You do not need a better tool for this. You need one pass that labels the guess before you quote anything.
- Keep the raw string. Save the transcript exactly as returned before you tidy it; once a casing is normalised, the earlier form is gone.
- List every capital on a name. Brand, person, place, product. Do not assume the capital means the spelling is correct.
- Replay around the name. Three to five seconds is enough. Decide the spelling from the audio and from any on-screen text, not from the capital.
- Check the on-screen branding. Captions and lower thirds were typed by a person; the printed form is stronger evidence than the recogniser's casing.
- Record the confirmed form. Write it beside the original so the next reader knows the capital was checked, not assumed.
One rule above all: if you cannot verify a capital on a name, leave it visible rather than silently “fixing” it. Silent normalisation makes the text 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 capitalisation layer applied to it. That is verifiable: the copy button reads the same element the download writes to a .txt file.
It is also narrower than it sounds. A raw word sequence is not a verbatim record. The model still guesses how to spell a name and still has no way to hear a capital, so the casing it emits is its own. A local generator removes one layer of writing decisions; it does not turn the remaining text into evidence. If you need publishable casing, you add it yourself — knowing you are the one adding it.
What you can check
Was the word spoken at all? Which of two near-sounding names? How is the brand spelled on screen? Did I keep the raw string?
What stays a writing decision
Which letters are capitalised Whether a brand is Instagram or instagram Whether It's or Its was meant Where a sentence starts with a capital
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 (capitalisation listed among display layers; inverse text normalization not user-configurable) and from a Chinese technical-community article on recognition error correction under multi-model fusion and mixed-language practice (named entities as a hard case for recognition, which capitalisation rides on). 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 capitalise words that should be lowercase?
- Capitalisation is a display layer that infers proper-noun status from context. When the inference is wrong, a common word gets a capital the audio never implied. It is a writing decision, not a sound.
- Why does my transcript capitalise a brand name wrong?
- The capitaliser guesses the casing of a name it may not even recognise, often from the first letter only. A brand can come back as “INST-Agram.” or “iNST agrom.” with no change in the audio. Check the on-screen branding.
- Can I turn off automatic capitalisation on a hosted service?
- Azure's display text format documentation states inverse text normalization is not user-configurable, and capitalisation is a separate display layer applied to results. Modes may change how the layer behaves, but you generally cannot request an unformatted word sequence.
- Is a capitalised name proof the spelling is correct?
- No. A capital is a claim about spelling that the audio did not make. “nor Easter” for “nor'easter” and “INST-Agram.” for “Instagram” both look deliberate and are both wrong.
- Does a browser-based generator avoid this?
- It removes the hosted display pipeline, so text arrives without a service-side capitalisation layer, but the local model still guesses name casing and cannot hear a capital. The verification pass is still required.
- What should I do before quoting a name from a transcript?
- Keep the raw string, list every capital on a name, replay the seconds around it, check on-screen branding, and record the confirmed form. If a capital cannot be verified, leave it visible.
How this connects to the rest of the site
Capitalisation is one writing decision among several. These pages cover the measurements and the neighbouring cases this page draws on.
- For the measured casing flips (
notebook→Notebook,INST-Agram.,It's→Its), see the eight audio cuts comparison. - For a different spelling guess the recording cannot settle, see why a transcript writes it's and never its.
- For the broader mechanism — punctuation, digits and filtering are written too — see why a transcript adds punctuation you never said.
- For the frame that ties these layers together, see why a transcript gives different text each time.
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.
Related case: why a transcript drops filler words — a display layer that subtracts tokens instead of rewriting them.