A longer file does not make you wait longer in a straight line, and it does not return more words in a straight line either. On the machine behind this page, a 39-second clip took 39.6 seconds and gave back 107 words; a 3-minute-9-second file took 92.9 seconds and gave back 537 words. The file got almost five times longer, but the wait only a little more than doubled. That is not a bug. It is what happens when a browser tool cuts audio into fixed-size pieces and pays most of its cost once, before the first piece is read.
This page explains the three things that change when a recording gets long: where the cut lands, how the wait grows, and why the same file can come back with three different word counts. All numbers below are runs that actually happened on a 2-core, 4 GB Windows 11 cloud machine running Chrome. None are estimates, and none are a promise about your device.
Where the cut lands decides the words you get
A speech model does not swallow a three-minute file in one bite. The audio is split into pieces of at most 30 seconds, and each piece is cut at the quietest moment near its mark. That sounds like a detail. It is the whole answer to a common complaint: two people feed the same Reel to the same tool and get different transcripts.
We tested this on one 3:09 file, three ways. A hard cut every 30 seconds returned 455 words. The library's own long-file mode returned 428 words. Cutting on silence returned 537 words. Same audio, same model, three different results — a spread of more than a hundred words, with nothing changed but the place where the pieces were divided.
The reason is mechanical. A cut that falls inside a word trims that word off one piece and may leave half of it on the next. A cut on silence leaves whole sentences intact. Neither result is more "true" than the other; they are different readings of the same sound. If a word looks chopped in half at the join between two passages, this is the first thing to check, before you decide the model misheard you.
The wait is not a straight line
Here are the four timed runs, from button press to finished text, on the same machine:
| File | Length | Size | Wait | Words out |
|---|---|---|---|---|
| MP4 720x1280, H.264 + AAC | 0:39 | 538 KB | 39.6 s | 107 |
| MP3 | 0:37 | 220 KB | 38.7 s | 107 |
| WAV 16-bit PCM | 0:37 | 1.7 MB | 37.4 s | 107 |
| MP4 720x1280, H.264 + AAC | 3:09 | 2.6 MB | 92.9 s | 537 |
The 0:39 clip ran at roughly real time. The 3:09 file ran at about 0.49x real time: it got 4.8 times longer, and the wait grew only 2.35 times. A fixed cost is paid once — loading the model, preparing the file, setting up the browser — and after that the extra audio is comparatively cheap. This is why a short clip can feel slow for its size, while a long one is the better deal per second of speech.
It also explains why a first run is slower than the ones after it. The model is downloaded once per browser: 43.5 MB in total, made up of 39.6 MB of model files and 3.9 MB of runtime. After that the page works offline. If your first file takes noticeably longer than the table suggests, you are paying that download, not a per-file cost.
What "unlimited" leaves out
Search for a transcript tool and you will meet the word unlimited a lot. One tool we read live advertises a concrete ceiling instead: a file up to 25 MiB and 5 minutes, processed in the browser. A page on that same site promises to explain where a five-minute cap comes from and how to split and stitch a longer file. Other tool pages we read use unlimited as a selling line and say nothing about what happens near the top of the range.
Neither framing tells you what the cut point does. A limit is a gate: above it, you cannot pass. A cut is a seam: below the limit, the file is still divided, and where the seam falls changes the words. You can be well inside a generous limit and still get a transcript that drops a word at each join. That is the part worth understanding, whether your tool caps at five minutes or promises no cap at all.
The ceiling we actually measured
The longest file we have tested is 3 minutes 9 seconds. Nothing longer was measured here, so nothing longer is promised here. That is a smaller number than several tools advertise, and it is stated plainly because a measured boundary is more useful than a large untested one. If your recording is longer than that, the honest options are to split it into files this tool has actually handled, or to use a service that has published a higher tested ceiling.
Splitting is not the same as cutting a single file inside the tool. When you supply two files, each one is prepared and read on its own, so each pays its own setup cost and each gets its own set of seams. Read the pieces back to back and you may still find a word missing at a boundary. Leave a breath of silence at the start and end of each piece when you record, and the cuts have somewhere harmless to land.
Three questions people ask about long files
Does a bigger file always cost more time? No. In our runs the wait grew more slowly than the file: 4.8 times the length, 2.35 times the wait. Setup cost dominates short clips and matters less as clips grow.
Why did the same file return a different word count? Because the cut point moved. A hard 30-second cut, the library's long-file mode, and a cut on silence gave 455, 428 and 537 words on one 3:09 file. Ask where the seams fell before you blame the model.
Can I feed it an hour-long recording? Not with any measurement we can stand behind. Our longest tested file is 3:09. Split it, or use a tool that has published a tested ceiling above that.
Before you send a long file
Check the length against something that was actually measured, not against a word like unlimited. If you are near or above three minutes, expect seams and expect the word count to move with them. Listen at the start and end of each piece, read the proper nouns and numbers against the audio, and keep the file you sent so you can re-run a section that reads badly. A transcript is a readable draft of what was said, not a verbatim legal record, and the seams are the first place to look when it does not match.