How transcription in your browser works, and why nothing is uploaded

"Free online transcription" almost always means: your recording is uploaded to a company's server, a model runs there, and you get text back. That works, but it has a cost that is easy to forget until the recording is a client call, a medical consultation or a lecture you were not supposed to share. On this site the transcription runs in the opposite direction: the model comes to you. Your browser downloads it once, and from then on the recording never leaves your device. This guide explains how that is possible, what it costs you in time, and where its limits are.

The model is open source, and small enough to download

The speech recognition itself is done by Whisper, a model OpenAI released in 2022 under an open licence. It was trained on hundreds of thousands of hours of speech in many languages, and it comes in several sizes. The three offered here are the ones that fit comfortably in a browser:

Setting	Model	One-time download	Typical speed
Fast	Whisper tiny	about 39 MB	2 to 4 minutes per 10 minutes of audio
Accurate	Whisper base	about 73 MB	4 to 8 minutes per 10 minutes of audio
Best	Whisper small	about 238 MB	25 to 50 minutes per 10 minutes of audio

Those downloads happen once. The browser keeps the model in its cache, so the second time you transcribe something, even weeks later, there is nothing to fetch. The speeds above were measured on the processor of an ordinary laptop (2021, Intel Core i5) and are slower on phones. On a computer with a graphics card that can run the model, a larger version of the same model is downloaded instead (roughly double the size) and runs on the graphics card. Older built-in Intel graphics (Iris Xe and earlier), common in laptops, were slower than the processor in our tests, so those computers use the processor. The choice is made automatically.

What happens when you drop a file

1. Audio is extracted. Video files carry the picture and the sound together. The browser loads FFmpeg, the open-source media engine, compiled to run inside the page (about 31 MB, also cached), and uses it to pull out just the audio track, converted to the 16 kHz mono format the model expects. A two-hour video becomes a few dozen megabytes of audio in memory.
2. The audio is cut into windows. Whisper listens in chunks of thirty seconds. The audio is split into overlapping windows so that a sentence crossing a boundary is not lost.
3. Each window is transcribed. This is the slow part. The model turns sound into a sequence of tokens, one small piece of a word at a time, and the result for each window is stitched onto the previous one, with timestamps.
4. You get text and timings. The full transcript appears once the last window is done, and you can copy it, download it as a text file, or (on the subtitle page) as an SRT or VTT file with the timestamps included.

At no point in that sequence is a network request made with your audio in it. You can verify this yourself: load the page, let the model download, switch your device to airplane mode, and transcribe. It still works.

Why it is slower than the upload services

A transcription company runs the model on racks of dedicated GPUs and gives each file a slice of that power. Here the model runs on whatever you are holding. A recent laptop with a graphics card handles the Accurate model at several times real time, so a 10-minute recording takes a couple of minutes. A mid-range phone running the same model on its main processor can take as long as the recording itself, or longer. The page shows which stage it is on so you can tell the difference between "still working" and "stuck", and you can cancel at any time.

Two practical consequences follow. First, keep the tab open and the screen on while it runs; a phone that locks its screen may pause the work. Second, for a long recording on a phone, use the trimmer to cut out the part you need, or do the long file on a computer.

Fast, Accurate or Best?

The three settings are the same model family at three sizes, and the differences are predictable. Fast is good for a quick draft of clear English speech and poor at everything else. Accurate is the sensible default for English, Spanish, Portuguese, French, German and the other large European languages. Best is noticeably better with accents, background noise, technical vocabulary and names, but about six times slower than Accurate, and it is the only setting that reliably handles several languages, among them Hindi, Tamil, Urdu, Turkish, Greek, Hebrew, Vietnamese, Thai and the Nordic languages. The page tells you when the language you picked needs Best, and offers to redo the file with it.

Every one of the 29 languages offered has been checked with a real recording, not a synthetic voice, and the page only lists a language if the model produced usable text for it. Languages where the model gave romanised or garbled output were left out rather than listed and left to disappoint you.

The exception: Instant mode

There is one way to send audio off your device on this site, and it is clearly labelled and off by default. Instant mode uses the speech recognition built into the browser itself, which in Chrome, Edge and Safari means the audio goes to that browser maker's speech service. It needs no download and works on weak phones, which is why it exists. The cost is privacy and precision: the recording is sent to a third party, the transcript arrives in real time (10 minutes of audio takes 10 minutes), and the timestamps are approximate. If you never switch it on, nothing ever leaves your device.

What to expect from the text

Whisper produces a draft, and a good one, but a draft. Expect to correct proper names, product names, numbers read out loud, and the occasional sentence where two people talked over each other. Punctuation is generally sensible; paragraphs are not, because the model does not know where your meeting changed topic. Recordings with a lot of hiss or fan noise transcribe worse, and the noise remover on this site can clean them first. Music, applause and silence are usually skipped, though a long stretch of noise can occasionally produce a hallucinated phrase, which is a known behaviour of the model and another reason to read the result before relying on it.