How it works

"Your audio never leaves your device" is an unusual claim, and you should not have to take it on faith. This page explains the whole pipeline in plain terms, names the parts that do talk to the network, and ends with a way to verify the claim yourself in about thirty seconds.

The short version

A speech recognition model is just a file full of numbers plus code that runs it. Normally that file lives on a company's server, so your audio has to travel to it. MeetingScribe inverts that: the model file travels to you, once, and then your audio is processed by your own processor or graphics card inside the browser tab. The audio never gets packaged into a network request, because there is nothing to send it to.

Step by step, in the tab

1. Reading the file

When you drop a file in, the browser reads it from disk into memory. Nothing is copied anywhere else. The page first probes the file for its duration, which is how it can warn you before you start a recording that is too long to finish comfortably.

2. Decoding and resampling

Audio files are compressed, and every format compresses differently. The browser's own audio engine decodes MP3, WAV, M4A, OGG, or FLAC into raw samples, then those samples are converted to a single channel at 16,000 samples per second, which is the format Whisper expects. This is also why AMR files fail: browsers have no decoder for them, so nothing on this page can help until you convert the file to MP3 or WAV first.

If you selected several files, this is the point where they are joined. Files with names ending in part1, part2, part3 are ordered numerically rather than alphabetically, so part10 lands after part9 instead of after part1.

3. Loading the model

The transcription engine and the model weights are fetched from a public content delivery network the first time you transcribe. This is the one large download, and it is the reason the first run is slower than every run after it. You choose the size:

• tiny, around 50 MB. Fastest, least accurate. Good for a quick check of a clear recording.
• base, around 90 MB. The default, and the right balance for most meeting audio.
• small, around 270 MB. Noticeably better on accents, crosstalk, and noise, at the cost of a bigger download and a slower run.

Once downloaded, the model is kept in your browser's cache. Later visits reuse it without downloading again, which is why the second transcription starts almost immediately. Clearing your browser's site data removes it and the next run downloads it again.

4. Transcribing

The model runs through ONNX Runtime, which uses your graphics card through WebGPU if your browser supports it, and falls back to WebAssembly on the processor if not. The page tells you which one you got in the line above the drop zone. WebGPU is dramatically faster, and is the main reason Chrome or Edge on a desktop outperforms an older browser or a phone.

Whisper does not read an hour of audio in one go. The audio is fed through in thirty second windows with a five second overlap between them, so words that fall on a boundary are not cut in half. The result is a list of text fragments, each with a start and end time. Those timestamps are what make subtitle export possible later.

To keep memory and speed reasonable in a browser, the model weights are quantized, meaning the numbers are stored at lower precision than the original research model. This is a real tradeoff: it is why in-browser transcription is fast enough to be practical, and also why a large server-side model will still beat it on difficult audio.

5. Separating speakers

Speaker separation, or diarization, is a second model that runs on the same audio. It does not recognize who anyone is. It divides the timeline into stretches and groups the stretches that sound like the same voice, producing anonymous labels like SPEAKER_00 and SPEAKER_01. You can then rename those labels to real names yourself, in the page.

Assigning speakers to words is a matching step: for each fragment of text, the page finds the speaker stretch that overlaps it most and attaches that label, then merges consecutive fragments from the same speaker into one paragraph. This works well with one to three voices. In a crowded meeting with people talking over each other, it will make mistakes, and if the diarization model fails entirely you still get a complete transcript without labels rather than an error.

6. Output

From the same result the page builds four things: a speaker-labeled transcript, a plain continuous transcript, an SRT subtitle file, and a VTT subtitle file. The transcript is editable in the page before you download it, which matters because no speech model is perfect and fixing three names is faster than re-running anything.

What does leave your device

Being precise about this is more useful than a blanket promise. Network requests this page makes:

• The page itself. HTML, styles, fonts, and scripts, like any website.
• The transcription engine and model weights. Downloaded from a public CDN on first use. This is a download to you, and carries no information about your audio.
• Analytics.A cookieless analytics script records that a page was viewed and that events like "transcription completed" happened. It records no file names, no audio, and no transcript text.
• Advertising. Only if you accept cookies in the consent banner. Decline and the advertising script is never loaded. Ads are how the site stays free. See the cookie policy.

What is never in that list: your audio file, your transcript, your subtitle files, your speaker names, or your edits. There is no server-side component that could receive them.

Verify it yourself

There are two easy checks, and you do not need to trust either our word or each other.

1. Watch the network.Open your browser's developer tools with F12, select the Network tab, then transcribe a file. You will see the model download and the usual page assets. You will not see an upload of your audio: sort by size and look at what is going out rather than coming in.
2. Pull the plug. Transcribe one short file so the model caches. Then disconnect from the internet entirely and transcribe another file. It still works. Nothing that requires a server can survive that test.

Why the forty minute guidance exists

This is the most common frustration, and the reason is memory, not an artificial limit. Decoded audio is uncompressed. A stereo recording at CD-quality sample rates uses roughly 350 KB for every second, so forty minutes of it occupies hundreds of megabytes in memory before the 16 kHz version, the model, and the intermediate results are added on top. A browser tab has a ceiling on how much it may hold, and crossing it kills the tab.

The practical workaround is to split long recordings into parts, name them in order, and select them together. They are joined into one transcript. Choosing the tiny or base model also leaves more headroom than small does.

The notes feature and where AI fits

After a transcription you can copy a ready-made prompt to your clipboard, containing your transcript plus instructions for writing a summary, key discussion points, and action items without inventing anything. That copy is a clipboard operation, and nothing is transmitted by the copy itself.

If you then paste it into an AI chat, that is your own action on a service you chose, and the text you paste goes to that service under its terms. We deliberately did not build a one-click cloud summary button, because it would send your transcript through a server and quietly undo the thing this tool exists to do. The about page explains that decision in more detail.

The honest caveats

• This design protects your audio from us and from third parties. It does not protect you from your own device. If your computer is compromised, nothing in a browser tab helps.
• We describe the tool as designed for confidential workflows. That is a description of the architecture, not a claim to any regulatory certification, and your compliance obligations remain yours.
• Accuracy depends on your audio. A clear single voice on a decent microphone transcribes very well. Four people on a laptop microphone in a room with air conditioning will not, with any tool.

Ready to try it? The transcriber is on the home page, free and without a signup, or read why on-device transcription matters if the privacy tradeoff is the part you are weighing.