Is there an Instagram transcript API?
Not from this site, and it is not being held back behind a sign-up. There is no server here to answer an API call: the whole site is 30 static files on a CDN, and the only code that transcribes anything is 22,349 bytes of JavaScript that runs in the visitor’s own tab. Measured on 10 November 2026, a POST to /api, /api/ or /api/transcribe on this site’s own Pages host returns 405 with a 0-byte body, and so does a PUT to /, /index.html or /robots.txt — the host accepts GET and HEAD and nothing else, because a static file has nothing to run. An API here would not be a feature that got switched off; it would be a different service, and the next section prices what it would have to hold.
What follows is the evidence: which lines of app.js a caller would have to reach and why a page cannot expose them, what the 405s and the CORS header on this host actually say, the memory and download arithmetic behind a hypothetical /v1/transcript, what the one site of the three that markets a transcription API says about it in its own words, and the two ways developers usually get round all of this without an API at all.
What a caller would need, and where the only code that does it lives
An API is a promise about two things: that a request arrives somewhere that will act on it, and that the acting code can be reached without a person in front of it. The first is answered by the 405s above. The second can be counted in the deployed file, and it is not there.
- Nothing in app.js is exported. The file is 22,349 bytes over 579 lines and contains 0 occurrences of
export, 0 ofwindow.and 0 ofglobalThis, so no function in it has a name an outside caller could hold. Even in a browser tab the transcription entry points are reachable only through the page’s own controls: 8addEventListenercalls, and every one of them is attached to an element inindex.html. - The one call that matters takes a decoded buffer, not a file. The recognition call is
mod.pipeline('automatic-speech-recognition', MODELS[key].id, …)at line 174, and the thing it is fed is the Float32 array thattoPcm()produces afterdecodeAudioDataand a resample to 16,000 Hz mono. A caller sending a URL or a file would have to reproduce that decode path first; there is no function that accepts a link and returns text. - There is no output a caller could read. The result is written into three elements on the page — the panel title, the
#metaline and#text— and copied or downloaded from there. No serialiser turns it into JSON: 0 occurrences ofFormData, 0 ofXMLHttpRequest, 0 ofWebSocket, 0 ofEventSource, 0 ofsendBeacon, 0 ofpostMessage. - The only two requests the page ever makes are downloads of its own machinery. One
import(for the recognition library and onefetch(for the model’sgeneration_config.json. Neither carries any part of the visitor’s file; a request that carried a file would have to be written, and there is nothing in 22,349 bytes that writes one.
What the host actually answers, and what it does not
Static hosting does not fail an API call by refusing it politely — it fails it by not having a method to route it to. These are the measured responses, taken from this machine on 10 November 2026 against getigtranscript.pages.dev with a desktop Chrome user agent.
getigtranscript.pages.dev on 10 November 2026: GET / 200, POST /api, /api/ and /api/transcribe 405 with a 0-byte body, OPTIONS / 405, and PUT /, /index.html, /robots.txt 405. The 405 rather than 404 is the detail worth keeping: the host is refusing a method, not failing to find a path, so no amount of URL guessing on any origin turns this into an endpoint. Access-Control-Allow-Origin: * appears on the GET and on the OPTIONS reply, which is ordinary static-hosting behaviour for a public file and says nothing about whether anything accepts a body.The CORS header is the part most likely to be misread. Access-Control-Allow-Origin: * was present on the GET and on the OPTIONS reply to /, and it was also present on the three 405 replies to PUT. A permissive origin header on a static site means the files are readable from anywhere; it does not mean a browser tab can push work at them. The method is what decides, and the method is the one thing this host will not accept with a body.
GET / and GET /robots.txt returned 200, while POST /api, POST /api/, POST /api/transcribe, OPTIONS / and PUT /, PUT /index.html, PUT /robots.txt all returned 405 with a 0-byte body — not 404, including on the three paths that do not exist. The header that misleads is Access-Control-Allow-Origin: *, which appeared on the 200 and on the OPTIONS reply and also on the three 405s: a permissive origin header says the files may be read from anywhere, and says nothing about whether a body will be accepted.What the same request would cost if there were a route behind it
Suppose the 405 were replaced by a function with a model attached. The bill is not a secret; it is arithmetic on numbers that are already in the deployed site and in the model repository. This is what a single request would hold.
- The weights are a fixed 76,894,612 bytes before any audio is read. In the repository,
onnx/encoder_model_int8.onnxis 23,201,314 bytes andonnx/decoder_model_merged_int8.onnxis 53,693,315, which sums to 76,894,629 — 17 bytes more than the figure the page quotes, because that figure is a rounded 76.9 MB. A caller pays this once per cold instance, not once per request, so it decides how many requests a warm instance can answer in its lifetime rather than the size of one. - Ten minutes of audio is 38,400,000 bytes of model input. The route would decode to 16,000 Hz mono, and a 32-bit float sample is 4 bytes, so one second is 64,000 bytes and 600 seconds is 38,400,000 — half of the 76.9 MB above. Twenty minutes is 76,800,000, which is 1.88 times the tiny checkpoint’s 40,844,231 and 0.9989 times the base checkpoint’s 76,894,612: the audio and the model become the same size at that length.
- Both buffers are live at once. The page keeps the decoded source and the 16 kHz render in memory simultaneously, which is why an hour of audio is 230,400,000 bytes of audio alone, and why a request had better carry a length cap rather than a timeout.
- The per-token arithmetic is 2,300,000 multiply-adds per second of audio. The encoder is an INT8 graph, so its parameter count is its byte count: 23,201,314 at 2 operations per multiply-add is 46,402,628 operations per input window, and a window is 30 seconds of audio, so 1,546,754 operations per second of audio — 1,546,754 per second, 92,805,234 per minute, 928,052,340 per ten minutes. This is a lower bound that ignores attention and the decoder; it is here to show the shape of the cost, not to predict a latency.
- Everything the page does per run, a server would do per request. The page serialises model loads through a queue precisely so two pipelines never run at once in one tab; a route would need the same discipline across requests, plus a limit on how many requests one instance accepts before it is recycled, because the 76.9 MB above is not shared between them.
What a transcript API actually looks like, in the words of one of the three sites people compare
One of the three best-known tools in this category markets a transcription API, and it is the only one of the three whose pages say anything about the subject at all. On wayin.ai’s general transcript tool page, the section headed “Integrate Video & Audio Transcription API” reads: “With WayinVideo Transcription API, you can seamlessly process all kinds of video and audio content at scale. Unlock accurate subtitles and speaker-labeled transcripts with just a simple API call. Reach a global audience instantly with built-in translation support for over 100 languages.” Its Instagram tool page carries “API” twice, both times in the site-wide navigation and footer rather than in the page’s own copy.
Across all 70 text files in this site’s benchmark directory — 616,517 characters, of which the three competitors’ Instagram pages account for 10 files and 126,491 characters — the uppercase word “API” appears 36 times, and every one of the 36 is on wayin.ai. speakapp.com: 0. getthescript.app: 0. Related vocabulary is just as absent: SDK 0 times, endpoint 0, webhook 0, developer 0, integrat 0, Zapier 0, automation 0. The word “batch” appears in exactly one of the three Instagram pages, in getthescript.app’s own tab row — “Single Video” and “Batch Mode” — which is a user interface for several files, not a programmatic interface for one.
Two routes that work without an API, and what each one costs
Almost everyone who searches for an Instagram transcript API does not want an endpoint. They want a hundred transcripts without a hundred exports, and there are two shapes of answer that need no server on this side. The first is a script of your own that drives a browser; the second is to give up on automation and let the tab do one file at a time. Both are honest about what they cost.
The cost of the first is the model download, and it is repeated. Each fresh browser profile or each cleared cache pulls the weights again: 76,894,612 bytes for base, 40,844,231 for tiny. At 30 seconds of audio per recognition window, 1,000 windows is 30,000 seconds of audio — and at 64,000 bytes a second that is 1,920,000,000 bytes of audio to arrive at one transcript of the same order of size as a download you already paid for once. A server-side queue does not change that ratio; it only decides who waits.
The cost of the second is your time, and the tool’s own numbers bound it. The model is fetched at the first file, not at page load, so a script that opens a new tab per item starts the 76.9 MB over again in every tab it opens, and one that reuses a single tab pays it once. That single sentence — one tab, many files — is worth more to someone processing a series than any wrapper written around this page, because there is no wrapper to write.
Why this page is not going to add one
An endpoint that accepts an Instagram link would also have to accept the server-side half of the problem: fetching the media from Instagram, which is the same request that returns a login wall to a machine, and then holding a decoded hour of audio at 230,400,000 bytes per concurrent request. The version of this tool that exists was built the other way round — the file stays in the tab, the model travels to it — and every number on this site is measured in that configuration. Publishing an API would not extend the tool; it would replace its one distinguishing property with the thing all three competitors already do.
If what you actually need is to get through a series of files, the bulk page is the honest description of what one tab can take, and the upload page has the four checks that show nothing left the machine. If the file never existed as a file, the link page explains why a link is not something this tab can open, and the no-login page covers the account question that usually sits behind an API search.
Last checked: 10 November 2026. Measured on this machine against this site’s own Pages host with a desktop Chrome user agent: GET / returned 200; POST to /api, /api/ and /api/transcribe returned 405 with a 0-byte body; OPTIONS to / returned 405; PUT to /, /index.html and /robots.txt returned 405; Access-Control-Allow-Origin: * was present on the GET and on the OPTIONS reply. The source counts are from the deployed files as they stand on this machine: app.js 22,349 bytes over 579 lines, with export, window., globalThis, FormData, XMLHttpRequest, WebSocket, EventSource, sendBeacon and postMessage each 0 times, addEventListener 8 times, import( 1 time and fetch( 1 time; the pipeline call is at line 174. The file sizes for the weights are the repository’s own listing for onnx-community/whisper-base and onnx-community/whisper-tiny, read on 10 November 2026. The byte and operation figures are arithmetic on RATE = 16000, one channel and a 32-bit float sample, printed and re-checked here before being written down; the 2 operations per multiply-add and the assumption that the encoder dominates latency are stated as assumptions, not measurements. The competitor counts cover 70 text files and 616,517 characters in this site’s benchmark directory, with the quoted sentences from wayin.ai reproduced exactly as they appear there.