Instagram transcript summarizer: why the summary is a second pass
There is no summariser here, and the reason is structural rather than commercial. The thing that turns speech into text in this tab is a speech recogniser, and 22,349 bytes of app.js contain 0 occurrences of chunk_callback, 0 of max_new_tokens, 0 of temperature and 0 of the word “summar”: there is no second model in the page and no setting that would make the first one do a second job. What the page hands you is the whole thing — measured on 27 October 2026, a 600.3-second clip came back as 1,194 words, 111 segments and 8,173 characters on screen, 6.85 characters for every word spoken, with nothing dropped. A summary is a different operation on top of that: it deletes words without replacing them, so it belongs after the transcript, in the hands of something that was built to write.
The rest of this page is the evidence: which lines turn audio into text and why none of them can compress, the arithmetic in characters and bytes that a summary has to work against, what a real summariser would add to a download that is already 43.6 MB, what the two of the three sites that sell this feature say their output actually is, and four ways to get a summary from a transcript you already hold.
What the transcript pass can do, and what it has no dial for
The recognition call is one line and it has four keys. pipe(pcm, { language, task: 'transcribe', return_timestamps: true, chunk_length_s: 30 }) at line 278 of a 579-line file — a language, a task pinned to transcribe, timestamps on, and a 30-second window. That is the entire surface. A max-token limit, a sampling temperature, a beam width, a no-repeat penalty: all four are absent, so there is no knob that would make the model say less.
- The task cannot be anything but transcribe.
task: 'transcribe'is a literal at line 280, not a value read from the page; the model file itself pins the decoder’s task slot withforced_decoder_idsset to[[1, null], [2, 50359]], and 50359 is the transcribe token. The summarise path that a Whisper checkpoint technically has is not wired to anything. - Nothing in the page would call the result a summary. The output panel is headed “Transcript” at line 439 (
$('out-title').textContent = 'Transcript'), the file is written astranscript.${kind}at line 520, and the copy button returnscurrentBody(), which is verbatim what is on screen. - Nothing is cut by length either. Fed 39,999 characters in a sandbox run on this machine on 12 November 2026,
render()put 39,999 characters on screen and printed “8000 words” in the status line — the page has no truncation step to inherit, which is why a summary has to be a deliberate deletion rather than a leftover. - Nothing is stored, so a second pass cannot be added later by this page.
localStorage,sessionStorage,indexedDB,cachesandserviceWorkereach appear 0 times in the 22,349 bytes. Close the tab and the transcript is gone unless you saved it; a summariser added to this page would have to work on the same visit, or ask you to paste the text back.
app.js measured on 12 November 2026: one recogniser, four keys on the call at line 278, no token limit, no temperature, no chunk_callback. The right box is what a summary needs, and it is absent from the file altogether — 0 occurrences of the word “summar” in 22,349 bytes. The arrow is the only place an outside tool can join in, and it carries the same 1,194 words and 8,173 characters the measured 600.3-second run produced.Why a summary has to delete, and what deleting costs in characters
Speech recognition is close to lossless; summarising is not, by definition. The measured run is the yardstick: 1,194 words over 600.3 seconds is 119.3 words a minute, and 8,173 characters of stamped text is 6.85 characters per word. A summary that shrinks that by any meaningful factor has to remove words, and once a word is removed from a transcript no later pass can tell you what it was.
- A quarter-length summary of the measured clip is about 298 words. 1,194 × 0.25 = 298.5, and at the same 6.85 characters per word that is roughly 2,043 characters — which is the entire value of the deletion: about 6,130 characters of the original are gone, and the summary cannot be expanded back into them. This is arithmetic on the measured run, not a measurement of any product’s output.
- In a sandbox run of the deployed
buildSrt()andbuildVtt()on this machine, 33 cues of 76 characters each became 2,508 characters of text, 3,035 bytes of stamped body, 3,653 bytes of.srtand 3,661 bytes of.vtt. Keeping every fourth word turned the same 528 words into 132 words and 659 characters, 21.7 per cent of the stamped body — and the timestamps had to go with it, because a stamp tells you where in the audio a line you can no longer read used to be. - The deletion is not something the page does even when asked indirectly. The only editing step in
render()removes bracketed annotations such as[Music]or(smooth music)before counting words (line 412) and does not touch the text on screen (line 421). That is the one deletion in the file, it is for word counting, and it leaves the transcript itself whole.
What a summariser would add to a download that is already 43.6 MB
The two checkpoints the page offers are 76,894,612 bytes for the better-accuracy model (79.7 MB as the page quotes it) and 43,649,619 bytes for the smaller one (43.6 MB), and both are already int8-quantised and running on WASM rather than on a graphics card. A summariser is a language model with its own weights, which on the same machine would arrive as a third download, not as an option inside the first.
- The audio is not the expensive part; the models are. 600.3 seconds at 16,000 samples a second is 9,604,800 samples, which at 4 bytes each is 38,419,200 bytes of model input — exactly half of the 76,894,612-byte checkpoint. A second model of any useful size would therefore be the largest single object in the tab, and there is no build of
app.jsthat loads two pipelines at once:loadModel()serialises loads through one promise queue on purpose, because two pipelines on a phone is how a tab runs out of memory. - The page already showed what a second concurrent download costs. A comment at line 569 records the measurement behind the current behaviour: loading the model on page load left both downloads running when the visitor changed their mind, and “a run that should have cost 44 MB cost 118 MB”. That is the arithmetic any summariser would inherit, in the same tab, before it reads a single word.
- The transcript is small enough to paste, which is the cheaper route than any of this. 8,173 characters for a ten-minute clip is about 1.4 pages of text; the whole summarising step can be done by pasting that into a chat model you already use, with the transcript as the only input and no second download at all.
What the two sites that sell this feature say it outputs
Three sites dominate this category, and two of them publish a summariser page for Instagram; the third does not have one in the benchmark set held here. Both pages agree on the shape of the output and disagree about the name of one of its parts, and neither one puts a number on how much shorter the summary is than the transcript.
- getthescript.app names four outputs, and calls the fourth verbatim. Its page says: “Our AI summarizer extracts and organizes Reel content into 4 clear, high-value formats.” The four are listed as “Executive Summary”, “Key Takeaways & Highlights”, “Timestamped Chapters” and “Full Transcript & Subtitles”, and the FAQ repeats them: “You receive a structured executive summary, bullet-point key takeaways, timestamped chapters, and the full verbatim transcript”. The same page puts the two jobs next to each other in a comparison table, with “Raw text only (no summary)” under “Generic Transcribers” and “Automatic structured summary” under its own column — which is the cleanest statement of the point on this page: a transcript is the input, not the output, of a summariser. Note that its four formats include the transcript itself, so the summary is shipped beside the lossless version rather than replacing it.
- wayin.ai names three and adds a mind map. Its page describes “a structured Reels summary, timestamped transcript, and visual mind map”, exportable “as TXT, Markdown, or subtitles (SRT/VTT)”. It also claims a capability this page structurally lacks: “This Instagram Video Summarizer goes beyond audio, handling visual-based analysis for silent Reels like sports highlights” — a silent Reel is exactly the case where an audio-only route returns “No speech found” and 0 words.
- Across 30 pages and 227,427 characters of the three sites’ own visible text, the word “summar” appears 286 times and not one of them is attached to a compression ratio or a word count. The related vocabulary is thin: “takeaway” 25 times, “executive summary” 6, “TL;DR” 2 (both on getthescript’s TikTok and YouTube summariser pages, not the Instagram one), “condens” 1. A reader comparing the two pages has no number by which to compare them.
Four ways to summarise a transcript you already hold
None of these needs a summariser inside the transcription tool, and each one keeps the full transcript as the fallback if the summary turns out to be the wrong shape for what you want.
- Keep the timestamped text as the summary’s index, not as the summary. A 600.3-second clip yields 111 segments, about 11.1 segments a minute, and each one is a sentence-length handle. Reading the first line of every tenth segment gives you a 15-line overview with the ability to jump back to any of them — the one thing a generated summary cannot offer, because it has dropped the words that sat between the lines it kept.
- Cut the transcript into the summaries you actually wanted. Since 111 segments is more than any single page of notes, split the file at a topic boundary and summarise each part separately; the page’s own export gives you the segment boundaries in the
.srtfor free, each cue timed to the millisecond. In the 33-cue sandbox run the.srtcame to 132 lines and 3,653 bytes, the.vtt3,661 bytes, and the difference between them was exactly 8 bytes — theWEBVTTheader and the blank line after it — so the two files carry the same 33 boundaries and choosing between them is a question of which player you are feeding. - Do not summarise the second half of a mixed file separately. A transcript whose language switches mid-file comes back with the wrong words in the part that was in the other language, and a summary cannot tell a misheard word from a word the speaker chose. The window arithmetic makes the damage easy to predict: a 161.2-second file is 8 windows of 30 seconds, 240 seconds of audio through the model for 161.2 seconds of recording, and the language switch lands inside whichever of those 8 windows contains it. Fix the transcript first — the 19 wrong words in a 338-word measured run are the sort of thing to check against — and only then summarise, using the accuracy page as the list of what to look for.
- Keep the summary and the transcript as two files with two names. Every export from this page is called
transcript.txt,transcript.srtortranscript.vtt(line 520); a summary you write next to it will overwrite the transcript in the same download folder unless you rename one of them before saving the second.
The question behind the query
Someone searching for an Instagram transcript summarizer is usually not looking for a feature list; they are looking at a Reel that is too long to read. The honest answer to that is the one this page can measure: the transcript of a ten-minute Reel is 8,173 characters and arrives in about 0.42 seconds of processing for every second of audio, and the summary is a separate, lossy step that belongs wherever you already write. The tool that hears and the tool that shortens are two tools, and keeping them apart is what lets you reread the first one when the second one gets it wrong.
Last checked: 12 November 2026. Measured on this machine (Windows 11, one AMD EPYC 9754 core with 2 threads and 3.9 GB of RAM) unless marked otherwise: the 600.3-second run — 1,194 words, 111 segments, 8,173 characters on screen, 251.5 seconds of recognition inside 271.9 seconds end to end, 180 MB of JavaScript heap — is the run logged on 27 October 2026, and the 39,999-character render test and the buildSrt()/buildVtt() byte counts come from a sandbox run of the deployed app.js on 12 November 2026 (22,349 bytes, 579 lines, md5 0431b8cb5a4e008607f720f78766258e). Everything else is arithmetic on constants in that file — RATE = 16000 at line 236, task: 'transcribe' at line 280, chunk_length_s: 30 at line 282 — or on the model’s own generation_config.json, read on the same day: 1,194 ÷ 600.3 × 60 = 119.3 words a minute; 8,173 ÷ 1,194 = 6.85 characters a word; 600.3 × 16,000 × 4 = 38,419,200 bytes of model input, which is 0.4996 of the 76,894,612-byte checkpoint; 1,194 × 0.25 = 298.5 words. The checkpoint sizes are the repository’s own file listing for onnx-community/whisper-base and onnx-community/whisper-tiny. The competitor sentences are reproduced exactly as they appear in this site’s benchmark directory, which held 30 content pages totalling 227,427 characters of body text on 12 November 2026, with each page’s heading list stripped before counting. The 25 per cent compression and the 200-words-a-minute reading speed are illustrations on top of measured figures, not measurements of any product, and no search volume, difficulty or ranking estimate is given here because none was retrieved.
If what you want is the transcript before the summary, the accuracy page counts the wrong words in a measured run, the timestamps page explains the 15-character prefix that makes a summary jumpable, the SRT and VTT page covers the cue boundaries you can split on, and the repurposing page takes the transcription further into a draft than a summary does.