Why Does My Instagram Transcript Generator Pick the Wrong Homophone?
Because the microphone records a sound, not a spelling. When two words sound the same — its and it’s, their and there — the recogniser has one acoustic signal and must choose one written form, and it chooses the one its language model has seen more often. When a sound is only close, it can go further and return a different real word entirely: handwritten came back as “hundred and”, Nobody as “Everybody”. The result is never a typo. It is a legal English word that fits the sentence, which is exactly why a spellcheck, a grammar check and a quick read all let it through. This page explains where that choice is made, reports what this site’s own counted runs actually returned, and gives a one-minute worksheet for catching a swap before it reaches a caption.
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 choice happens at word selection, not at spelling. Chinese technical-community writeups on recognition errors place this at the seam between the acoustic model (what sound was heard) and the language model (which word sequence is likely): a writeup on multi-accent data alignment and acoustic versus language-model error sources describes the two disagreeing more often under accented, fast or mixed-language audio, and a second writeup on Cantonese and multi-model fusion describes recognition correction as picking a more probable word rather than reading one from the recording. So for a homophone the model resolves one sound to the higher-frequency form, and for a near-sound it resolves an uncertain candidate to a legal word that fits the context. This site’s own measured runs show both: across twenty-three public Reels the contraction it’s appeared 32 times and the possessive its 0 times, your 26 times against you’re once, and there 9 against their 3 and they’re 2; in a separate eight-cut experiment on a 182.961406-second synthetic narration, handwritten returned as “hundred and”, Nobody as “Everybody”, and play as place. In every case the output is a real word, so no automatic check flags it. The fix is to treat every homophone, name and number as a guess, run a substitution test, replay a few seconds, and check any on-screen text. Below: why the model picks one word, why the error survives proofreading, what the measured runs returned, and the worksheet.
One sound, several legal spellings
The waveform carries phonemes, pitch and timing. It carries no apostrophes, no distinction between their and there, and no marker that a sound belongs to two rather than to. A recogniser maps that signal to a sequence of words, and where the acoustic evidence is identical for two candidates, a language model breaks the tie using what it learned about how often each word appears in context. A Chinese technical-community article on multi-accent data alignment and the error sources of acoustic versus language models puts the disagreement between those two models at exactly this point, and notes the disagreement grows under accented or mixed-language audio — the normal condition for a fast Reel. A second Chinese writeup on Cantonese recognition and multi-model fusion describes the correction step as choosing a more probable word form rather than recovering a literal one.
That is why the result is a word, not a misspelling. The model is not guessing letters; it is choosing between two complete, correctly spelled entries in its vocabulary. There is nothing for a spellchecker to object to.
| In the audio | Chosen at word selection | Consequence |
|---|---|---|
| The sound /ɪts/ | it's or its |
Two legal words, one signal; frequency decides |
| The sound /jɔːr/ | your or you're |
Both grammatical in some sentences, so structure does not settle it |
| The sound /ðeər/ | there, their, they're |
Three legal words share one acoustic form |
| A near-sound, not an exact match | handwritten or hundred and |
A whole phrase can be swapped for another legal phrase |
The right column is a selection, not a transcription. Word choice sits at the acoustic/language-model seam described in the Chinese technical-community sources cited on this page.
Homophones and near-sound swaps are the same trap
It helps to separate two cases that behave the same way at the proofreading stage. An exact homophone gives the model no acoustic information at all about which spelling is meant, so the decision is pure language-model preference. A near-sound swap starts one step earlier: the acoustic model proposes an uncertain candidate and the language model resolves it to a legal word that fits the surrounding context. Both end at the same place — a valid word in a valid sentence that is not the word spoken.
The shared danger is that the output contains no non-word. A garbled token gets caught; a plausible replacement does not. The smoother the substituted sentence reads, the more likely it is to survive.
That smoothing is not hypothetical. In the eight-cut experiment the substitution Nobody → Everybody reverses the meaning of a sentence while leaving its grammar flawless, and handwritten → hundred and turns one adjective into a number plus a conjunction. Neither looks like damage. A reader who did not hear the audio has no reason to stop.
Why a spellcheck, a grammar check and a quick read all miss it
A spellchecker compares tokens against a dictionary; both candidates are in it. A grammar checker tests structure; there for their and to for two are grammatical in many frames, and only some substitutions (like it’s before a bare noun) produce a structural wobble. Reading the sentence out loud is worse, not better: it reproduces the same sound that caused the confusion in the first place, so the wrong word sounds right.
The one check that can settle it is a comparison against evidence outside the text — the audio itself, or typed on-screen words such as captions, lower thirds and brand names. Everything that stays inside the text is working from the same guess the model already made.
What this site’s own counted runs returned
Two local experiments published on this site, run with CPU speech models and no hand correction, show the choice landing on the more frequent form and drifting between runs. They are not Instagram data and do not measure any hosted service; they are here because the counts are re-checkable.
| Measurement | What came back | What it shows |
|---|---|---|
| 23 public Reels, exact-string count | it's 32 (100%), its 0 (0%) |
One sound resolved to the frequent form every time |
| Same 23-Reel sample | your 26 vs you're 1 |
The bias is strong but not absolute |
| Same 23-Reel sample | there 9 vs their 3 vs they're 2 |
Three-way ties still settle on the frequent form |
| Same 23-Reel sample | to hundreds vs too 1 and two 4 |
Function words dominate; rare forms nearly vanish |
| 8 cuts, 182.961406 s narration, 24 calls | handwritten → hundred and |
A near-sound resolved to a different legal phrase |
| Same eight cuts | Nobody → Everybody; play → place; and checked → and... check |
Swaps stay grammatical and readable |
| Same eight cuts, whole-file output | 435 words, endpoint 183.0 s against a 182.961 s input | The same configuration also overshoots its own duration |
| Same eight cuts, split outputs | It's → Its, 94 → 1994 |
The apostrophe direction reverses under split input |
Sources: the site’s 23-Reel it's / its count (faster-whisper base, int8, two threads, beam 5; two clips read with base, three with tiny after a memory fallback; exact strings only, sentence-initial capitals included, hyphenated compounds and partial matches excluded) and the eight audio cuts comparison (182.961406-second synthetic English narration, cached faster-whisper tiny, CPU int8, one thread, one worker, beam size one, English selected, VAD disabled, 24 calls across eight whole-file runs and sixteen halves). The differences are recorded as differences: neither run used an independently listened reference transcript, so they establish that the chosen word moves, not that any specific form was “correct.”
One detail is worth keeping rather than smoothing over. The 23-Reel count has it’s at 32 and its at 0, while the eight-cut split outputs contain It’s → Its in the opposite direction. These are two different experiments with different engines and different inputs, not a contradiction to be averaged away. The fact that the same sound resolves differently under different configurations is itself the point: the choice is a probability judgement, and probabilities move with the setup. Treating either run as a universal rule would be over-reading the data.
The quiet part: a swap often reads better than the original
A wrong word usually looks wrong. A substituted homophone does not — it is a word a person could plausibly have written, and because the model is biased toward frequent, well-formed written English, the replacement is sometimes tidier than the literal transcript. That is the trap. The more natural the sentence reads, the less likely a proofreader is to replay the audio, and the more likely the substitution reaches a published quote, a caption file or a set of show notes unchanged.
The rule is simple: any homophone, name or number in a transcript is a claim about spelling the audio did not make. It deserves the same replay-and-check as an obviously suspicious word, and more, because nothing about it looks suspicious.
A worksheet: how to catch a swapped word
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 homophone is normalised, the earlier form is gone.
- List the high-risk words. Scan for
it's/its,your/you're,there/their/they're,to/too/two, plus every name, brand and number. - Run a substitution test. Read
it'sas “it is”; if the sentence breaks, the possessiveitswas meant. Trytheirwhere you seethere. This catches the cases where grammar does expose the swap. - Replay around the word. Three to five seconds is enough. Decide from the audio, not from how clean the sentence looks.
- Check on-screen text. Captions, lower thirds and printed brand names were typed by a person; that printed form is stronger evidence than the recogniser’s choice.
- Record the confirmed form. Write it beside the original so the next reader knows the word was checked, not assumed. If it cannot be verified, leave it visible rather than silently “fixing” it.
Silent normalisation is the one move that makes this worse: it turns a visible guess into a confident-looking text that nobody can audit afterwards.
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 correction or fusion pipeline in that path, so the output does not arrive with a service-side layer re-choosing low-frequency words into high-frequency ones. That is verifiable: the copy button reads the same element the download writes to a .txt file.
It is also narrower than it sounds. The local model still resolves one sound to one word by probability and still cannot hear an apostrophe, so its can still come back as it’s. A local generator removes one layer of word selection; it does not turn the remaining text into evidence. If you need the exact word that was spoken, you add that verification yourself — knowing you are the one adding it.
What you can check
Was the word spoken at all? Which homophone fits the sentence? How is the name spelled on screen? Did I keep the raw string?
What stays a choice
its or it's their, there or they're to, too or two A near-sound's exact wording
The left column lists checks a person can finish against the audio or the frame. The right column lists distinctions no recording contains.
Scope and limits
The mechanism described here comes from two Chinese technical-community writeups on recognition error sources: one on multi-accent data alignment and the disagreement between acoustic and language models, the other on Cantonese recognition, multi-model fusion and correction as a choice among probable word forms. 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 write a different word than the one I said?
- The recording holds a sound, not a spelling. Where two words sound alike, a language model picks the more frequent written form, and where a sound is only close, it resolves an uncertain candidate to a legal word that fits the context. The result is a real word, so nothing looks misspelled.
- Why does it always write it's and never its?
- In this site’s count across twenty-three public Reels, the contraction it's appeared 32 times and the possessive its 0 times. The acoustic signal is identical for both, so the model resolves it to the form it has seen far more often. That is a frequency effect, not a grammar rule.
- Can a spellcheck or grammar check catch this?
- Usually no. Both candidate words are in the dictionary, and many substitutions stay grammatical. Only some frames, such as it's directly before a bare noun, expose the swap structurally. A substitution test against the audio is the reliable check.
- What are near-sound swaps, and how are they different from homophones?
- A homophone is an exact sound match with two spellings; a near-sound swap starts when the acoustic model proposes an uncertain candidate and the language model resolves it to a different legal word. This site’s eight-cut runs returned handwritten as “hundred and” and Nobody as “Everybody”. Both end as valid words that are not what was said.
- Can I turn off this word choice on a hosted service?
- Word selection is part of how a transcript is formed, not a switch most services expose. Correction and fusion layers exist precisely to choose a more probable word. Modes may change how the layer behaves, but you generally cannot request raw acoustic candidates instead of words.
- Does a higher confidence score mean the word is right?
- No. A word-level confidence score reports how sure the model is about a candidate, not whether that candidate matches what was spoken. The model is often most confident on the frequent form, which is exactly the one it substitutes.
- What should I do before quoting a word from a transcript?
- Keep the raw string, list high-risk homophones and names, run a substitution test, replay three to five seconds, check on-screen text, and record the confirmed form. If a word cannot be verified, leave it visible rather than silently correcting it.
How this connects to the rest of the site
Word choice is one writing decision among several. These three pages carry the measurements and the neighbouring cases this page draws on.
- For the 23-Reel count (
it's32 /its0) and the wider homophone table, see why a transcript writes it's and never its. - For the near-sound swaps (
handwritten→hundred and,Nobody→Everybody, 24 calls), see the whole-file and split comparison. - For the sibling layer that also writes capitals and punctuation after recognition, see why a transcript adds punctuation you never said.
- For the frame that ties the layers together — recognition, normalized, filtered, and re-run drift — 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.