Why Does My Instagram Transcript Generator Split Words Apart?
Because the audio has no word boundaries in it. A recogniser turns sound into a stream of phonemes, then a language model decides where one “word” ends and the next begins — that decision is spelling, not listening. So when a compound like nor’easter comes back as “nor” and “Easter”, or a time like 2.30 comes back as “2 .30”, nothing in the recording changed. The model placed a space where a human would not, or left one out where a human would put it. The danger is that a split word looks like a clean edit: “note book” reads like someone tidying the text, so it survives proofreading. This page shows why the boundary is a guess, what this site’s own repeated runs actually returned, and a one-minute worksheet to catch split words before they reach 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: word boundaries are a writing decision, not a sound. Technical writeups on recognition errors place word segmentation at the seam between the acoustic model (what was heard) and the language model (what word sequence is likely); the language model has no audio signal for “this is one word” — it picks the sequence it judges most probable, and that includes where the spaces go (a Chinese technical-community article on multi-accent data alignment and acoustic versus language-model error sources describes segmentation this way; a second writeup frames text formation as a sequence-labeling problem, i.e. a model decides where tokens begin and end rather than reading it from the audio). This site’s own measured runs show boundaries moving with no change in the audio: in five public Reels (353.1 seconds, 1,114 words) the compound nor’easter split into “nor” and “Easter”, and all three time expressions in one clip split into three tokens each (“five o ’clock”, “2 .30”, “7 .30”); in a separate eight-cut experiment a synthetic narration returned “note book” for “notebook” and the same stray-spacing pattern on figures. None of those spaces were in the recording. The fix is to treat every word boundary on a name, number or compound as a guess, scan for stray spaces inside words and figures, replay the seconds around it, and check any on-screen text. Below: why the model guesses the boundary, what the measured runs returned, why a split word is the quietest kind of error, and a worksheet you can run in a minute.
A word boundary is a writing decision, not a sound
The waveform carries phonemes, pitch, loudness and timing. It carries no spaces. A “word” is a unit a text system imposes after decoding, not a thing the microphone records. So word boundaries have to come from a step that reads the chosen phoneme or word sequence and decides segmentation — a language model, not the microphone. A Chinese technical-community writeup on recognition errors under multi-accent data alignment places segmentation exactly at the seam between the acoustic model (what was heard) and the language model (what sequence of words is likely), and notes that the boundary between “note” and “book” is the language model’s call. A second Chinese technical writeup frames inverse text normalization and similar text formation as a sequence-labeling problem — in other words, the text form is produced by a model that labels where tokens begin and end, not read from the audio.
That is why a hyphen or a space is written after decoding, by a layer you generally cannot switch off. English compounds — notebook, nor’easter — are exactly where this bites, because whether “notebook” is one word or “note book” two is a spelling convention, not a sound. The audio gives “note…book” either way; only the writing step decides if there is a space.
| In the audio | Added by the segmentation layer | Consequence |
|---|---|---|
| A pause between clauses | A space after a sentence boundary | The pause is real; the space is a writing rule |
| The sounds “note” then “book” | notebook or note book |
One lexical unit, two possible spacings |
| The figure “2.30” spoken as a time | 2.30 or 2 .30 |
The stray space is a segmentation guess |
| The storm name “nor’easter” | nor'easter or nor Easter |
One word, split into two by a boundary guess |
The right column is written, not heard. Segmentation sits at the acoustic/language-model seam and is described that way in the Chinese technical-community sources cited on this page.
Why the model guesses the boundary
A recogniser has no dictionary flag marked “join me” or “split me.” It returns a word sequence, and the segmenter decides token edges from context — the words before and after. That context-based guess is exactly where it can go wrong: a low-frequency compound the model does not recognise gets split at the most likely acoustic boundary, or two common words get joined into one because the language model expected a compound there. The same technical writeup notes that under multi-accent and mixed-language audio the segmentation error sources multiply, because the acoustic model and the language model disagree more often about where words are — which is the usual condition for Reels, often accented, fast, or switching languages mid-clip.
None of this is a “bug.” The segmenter is doing its job: turning speech into readable text. The problem is only that its spacing 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 boundaries 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 |
|---|---|---|
| Five public Reels, 353.1 s, 1,114 words (site measurement) | nor'easter → nor (11.84 s, conf 0.18) + Easter (12.00 s, conf 0.37) |
Compound split at the first break in that clip, 12% in |
| Same five-Reel sample, 67.9 s interview clip | All three time expressions split into three tokens: five o 'clock, 2 .30, 7 .30 |
Stray spaces where a human would write one word or one figure |
| Same five-Reel sample, tokens containing a digit | 6 tokens, 1 under 0.6 confidence (16.7%) | Numbers and names pool the doubt — and the boundaries around them drift |
| Eight audio cuts, 182.961406 s narration, 24 recognizer calls | notebook → note book |
A compound split by a space the audio never contained |
| Same eight cuts | notebook → Notebook; figure spacing like 2 .30 |
Casing and segmentation are both written decisions on the same word |
Sources: the site’s what breaks first (353.1 seconds, 1,114 words, five public Reels) and eight audio cuts comparison (182.961406-second synthetic narration, 24 calls, cached faster-whisper tiny, CPU int8, English selected). The boundary differences are recorded as differences; neither run used an independently listened reference, so they establish that boundaries move, not that any specific word was “wrong.”
The dangerous part: a split word looks like a tidy edit
A wrong word usually looks wrong. But “note book” for “notebook” reads like a careful writer who simply spaced a compound — it looks intentional, so it survives proofreading that would catch an obviously wrong word. The nor’easter case is the same: “nor Easter” reads like two words, both capitalised, as though it were a person’s name, when it is one storm. This is why word-boundary errors are among the most likely to reach a published caption: they look like formatting, not mistakes. The rule is simple — any space inside a name, number or compound is a claim about spelling the audio did not make, and it deserves the same replay-and-check as a suspicious word.
Mixed language makes the guess shakier
Word-boundary rules are language-specific. A language that joins compounds English would split, or splits compounds English would join, gives the segmenter 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 segmentation layer inherits the same ambiguity. If the Reel switches languages around a compound, treat the boundary as doubly uncertain.
A worksheet: how to catch a split 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 space is normalised, the earlier form is gone.
- Scan for stray spaces. Look inside names, numbers and compounds:
2 .30will not match2.30;note bookwill not matchnotebook. A scan for spaces inside figures and names catches most seams in seconds. - Replay around the word. Three to five seconds is enough. Decide the boundary from the audio and from any on-screen text, not from how clean the spacing looks.
- Check the on-screen text. Captions and lower thirds were typed by a person; the printed form is stronger evidence than the recogniser’s spacing.
- Record the confirmed form. Write it beside the original so the next reader knows the boundary was checked, not assumed.
One rule above all: if you cannot verify a boundary, 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 segmentation layer rewriting spaces. 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 where words begin and end and still has no way to hear a boundary, so the spacing 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 spacing, 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 compounds? How is the brand spelled on screen? Did I keep the raw string?
What stays a writing decision
Where the spaces go Whether notebook or note book Whether 2.30 or 2 .30 Whether a hyphen is present
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 two Chinese technical-community writeups on recognition error sources: one placing word segmentation at the acoustic/language-model seam and noting that multi-accent and mixed-language audio multiply those error sources, the other framing text formation as a sequence-labeling problem. 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 split words apart?
- Word boundaries are a writing decision made by a language model after recognition, not a sound in the audio. When the model is unsure where a word ends, it can place or omit a space, so “notebook” becomes “note book” or “2.30” becomes “2 .30” with nothing in the recording changing.
- Why does my transcript split a compound like nor'easter?
- A compound is one lexical unit in writing but several sounds in speech. The segmenter guesses the boundary from context and can split it at the most likely acoustic edge — “nor'easter” came back as “nor” and “Easter” in this site’s five-Reel measurement, with no space in the audio.
- Why does my transcript break numbers like 2.30 into 2 .30?
- The figure “2.30” is one token to a human, but the model can return it as separate pieces with a stray space, because the boundary between digits is also a writing decision. This site’s measurement found all three time expressions in one clip split into three tokens each.
- Can I turn off automatic word segmentation on a hosted service?
- Segmentation is part of how the text is formed after recognition; it is not a switch you control on most hosted services. Modes may change how the layer behaves, but you generally cannot request an unsegmented phoneme string.
- Is a spaced word proof the spelling is correct?
- No. A space is a claim about wordhood the audio did not make. “note book” for “notebook” and “nor Easter” for “nor'easter” 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 segmentation layer rewriting spaces, but the local model still guesses word boundaries and cannot hear them. The verification pass is still required.
- What should I do before quoting a compound from a transcript?
- Keep the raw string, scan for stray spaces inside names, numbers and compounds, replay the seconds around it, check on-screen text, and record the confirmed form. If a boundary cannot be verified, leave it visible.
How this connects to the rest of the site
Word boundaries are one writing decision among several. These pages cover the measurements and the neighbouring cases this page draws on.
- For the measured splits (
nor'easter→nor/Easter,2 .30,notebook→note book), see what breaks first and the eight audio cuts comparison. - For the broader layer that also writes capitals and punctuation, see why a transcript adds punctuation you never said and why a transcript capitalises the wrong words.
- For segmentation under mixed-language audio, see mixed-language Reels.
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.
For a mark that is inferred rather than heard, see why saying “comma” writes the word, not the symbol.