A public Instagram Reel becomes text only when something transcribes its sound track, because the platform exposes no text layer that a third party can copy, and whether a pasted link produces anything at all depends on four conditions that the pages offering this service rarely state together: the post must be reachable without a login, the address must name a video post rather than a Story or a profile, the clip must contain speech, and its running time must sit inside the limit of the service you chose. When one of those four is false the result is an error, a spinner that never finishes, or an empty box, and the reader is left submitting the same address again instead of changing the condition that failed.
- Copy the address from the Reel's share menu rather than from the address bar of a profile or a saved collection.
- Open that address in a browser window where you are signed out, and treat a page that will not play there as unreachable for any third-party tool as well.
- Read the address itself and confirm it contains a Reel, post, or TV path rather than a Stories path.
- Play the clip with sound on and check that a person is speaking, because music, words drawn on the picture, and silent product shots contain nothing to transcribe.
- Note how long the clip runs and compare that against the length the tool you picked says it accepts.
- Paste the link once, and if the output is empty, change one of the four conditions rather than submitting the same address again.
- If the post cannot be reached without a login, get the video onto your device as a file and transcribe the file instead of the address.
The four boundaries, in the order they fail
Public does not mean visible to you. A clip you can watch while signed in, or one shared with a close-friends list, is still behind a login from the outside. Services do say they handle only public posts, but that sentence usually sits in a question near the bottom of the page, long after the input box has invited you to paste anything at all. Testing the address in a signed-out window first separates a link problem from a tool problem before you spend a run on it.
The shape of the address matters more than the app you copied it from. A link to a Reel, a feed video, or a live replay saved to a profile identifies a video object; a link to a Story, a highlight, or a profile page does not. Here the leading pages flatly disagree: one says a pasted link works for Stories, another says a Story cannot be fetched from a link at all and has to be uploaded as a file on a paid plan, and a third lists Stories among the formats it handles without explaining either route. If the path is a Stories path, save the clip as a file first, because that is the only version all three descriptions can agree on.
Speech has to exist in the audio, not just on the screen. A Reel can carry words as styled text over the picture, as the typed description under the post, and as a spoken voice, and only the third of those can be turned into a transcript. Pages that advertise working without captions are describing their own method, not the clip; a video with music and no voice returns nothing on every one of them. Words drawn on the frames can be read off with optical character recognition, but that yields fragments with no speaker and no clock.
Length is the boundary that produces the most confusing failure, because the clip is fine and the link is fine. The leading pages set their ceilings at wildly different points, from roughly the length of a single short clip up to recordings long enough to require a sign-in, while all of them describe themselves as handling any public Reel. Some add a recurring free allowance on top of the ceiling, so a clip well inside the stated limit can still be refused because the allowance has gone. The clip did not change; the boundary on the service side did.
What the leading pages cannot agree on
Read the results for this phrase side by side and their stated limits pull in different directions. One page accepts only short clips, another extends to lengthier ones with a sign-in, and a third sets a ceiling far above either. Language coverage is quoted at different sizes on different pages, and those pages do not even agree on what the figure describes: some say a clip recorded in another language comes back in that language and can be translated afterwards as a separate step, while others present language support as one flat number with automatic detection and no translation step at all. Accuracy claims appear as a single figure from an internal test on one page and as a wide range on another, with no shared set of clips and no shared way of counting a wrong word, so they cannot be compared with one another.
None of that is a reason to distrust the tools. It is a reason to stop treating the claim that any public video works as a description of the platform. It describes one provider's pipeline, which has its own length ceiling, its own free allowance, and its own list of address shapes it knows how to fetch. A link that fails on one service is not evidence that the Reel cannot be transcribed; it is usually evidence that you crossed one of that service's boundaries.
The same confusion shows up in the small print about exports. Several pages offer a plain text download and a timed caption file as though they were the same button, then attach the timed option only to clips past a length threshold or to a paid tier. A reader who needs the words on a clock can therefore paste a perfectly good link, receive prose, and conclude that the tool failed, when in fact it delivered exactly the one output it was willing to give for that clip.
Where the fetching actually happens
A plain web page cannot pull a video from another domain by itself, so a box that accepts a link necessarily relies on a server somewhere to go and get the media. The leading pages describe that step in incompatible ways: one says processing happens in real time in your browser with nothing kept on a server, another says the audio is transcribed server-side, and a third says the file is pulled and discarded once the text is on screen. Those are three different answers to the same question, and the difference is not about speed. It decides whose machine holds the audio while the words are being produced, which is a privacy question rather than a performance one.
If that matters to you, the alternative is to remove the fetch entirely. Save the clip onto your device and hand the file to a page that decodes it in the browser's memory; there is no address to resolve and no server in the middle. The guide on saving a Reel for transcription covers the ways to get that file, and a browser-local run needs the model files once before it keeps working without a connection. The trade-off is plain: you do the fetching yourself, and the tool never sees a URL.
What comes back, and the container question
What a transcription run returns is plain prose: the spoken words in order, with punctuation guessed by the model. Quoting, searching, summarising, and drafting all work from that. If you need words tied to moments on a timeline, the format guide explains the two containers, and plain text will not turn into one of them by itself.
Check the words before you use them
A result that appears is not automatically a good one. Names, numbers, brand terms, and words read aloud from a hashtag are the usual places a model goes wrong, and music sitting under quiet speech makes all of them worse. Read the whole text once against the clip and fix those before you quote anything. The troubleshooting guide covers the common failure shapes, the language limits page explains what an English-only model does with a clip recorded in another language, and the invented-words note describes the rarer case where a confident word was never spoken at all.
It is also worth knowing which words you will not find. The description the creator typed under the post is ordinary selectable text, but it is a summary, a credits list, or a punchline as often as it is a record of what was said. Text burned into the picture can be read by a person and cannot be copied cleanly by a machine. Neither is part of the audio transcript, so a reader who wanted either one will not get it from a link, however many times they paste it.
The short version
So the practical order is: check the address outside a login, confirm it names a video, confirm somebody speaks, check the length, and only then paste. When one of those fails, switching services rarely helps, because the leading tools draw their boundaries in the same four places and merely attach different numbers to them. When all four pass and you would still rather keep the clip off somebody else's machine, the file route is the one that takes the link out of the question altogether.