Transcribe audio and video to text, right on your device
Drop a recording and read what was said. Fix the words while you listen, then save subtitles or a document. Nothing is sent to a server.
- Audio sent to a server
- None
- Account and minute limit
- None
- Model download, once
- 44 MB · 80 MB · 252 MB
- Languages
- 15 listed, or detected automatically
- Saves as
- SRT, VTT, TXT, DOCX
Drop an audio or video file here
MP3, WAV, M4A, OGG, Opus, FLAC, WebM, MP4, MOV
Pick the tool for the job
Every page opens the same studio, set up for one kind of file.
Transcribe a recording
Subtitles and dictation
How it works
From a recording to a file you can use, without leaving this page.
meeting.m4a
Open a recording
Drop an audio or video file on the page. Your browser decodes it and draws the waveform; nothing is sent anywhere.
Press Transcribe
Pick a model and, if you know it, the spoken language. The first run downloads the model (44 MB for the small one) and keeps it for next time. Lines appear as the model works through the recording, thirty seconds at a time.
- 00:0400:09SRTVTTTXTDOCX
Correct and export
Select a line’s time to hear it, type over what the model got wrong, adjust where the line starts and ends, name the speakers. Save the result as subtitles or as a document.
What “on your device” means here
Cloud transcription services receive your file, process it on their servers and send the text back. Here the model comes to the file instead: it is downloaded to your browser and runs there.
One request does leave your device: the download of the model from Hugging Face, which sees your IP address like any website you visit. It receives no audio and no text. After that first download, transcribing makes no network request at all.
The exception is live dictation, which relies on your browser’s own speech service. That page explains where the audio goes and offers a local alternative. The full account is in the privacy policy.
Your device
Choosing a model
All three are OpenAI’s Whisper, a multilingual speech model, in versions small enough to run on a laptop’s processor. A larger model hears difficult audio better and takes longer.
| Model | Download, once | 45 s of speech takes* | Use it for |
|---|---|---|---|
| Small | Download, once: 44 MB | 45 s of speech takes*: about 3 s | Use it for: Clear speech, one voice, a first draft |
| Medium | Download, once: 80 MB | 45 s of speech takes*: about 15 s | Use it for: Languages other than English, accents kept right |
| Large | Download, once: 252 MB | 45 s of speech takes*: about 35 s | Use it for: Interviews, noise, names, anything you will publish |
* Measured on 3 October 2026 in Chromium on a desktop with 20 processor threads, with the model already downloaded. A phone or an older laptop will be several times slower.
Why use Transcribe Tools
- The file stays on your device
- Your browser decodes the recording and a Whisper model transcribes it in a Web Worker. Neither the audio nor the text is sent to this site.
- You see the download before it starts
- The speech model is the only download: 44 MB, 80 MB or 252 MB, depending on the one you pick. It is fetched once from Hugging Face and kept in the browser.
- An editor tied to the waveform
- Every line is a band under the waveform. Select its time to hear it, type over the text, and move its start and end.
- SRT, VTT, TXT and DOCX
- The four files are written in the browser from the same transcript, with speaker names when you add them.
- A reload does not lose your work
- The transcript is saved in this browser as you type, and “Clear transcript” removes it.
- Dictation says where your voice goes
- Live dictation uses your browser’s speech service and the page tells you so. Recording on the device is offered next to it.
Questions and answers
Is my audio uploaded to a server?
No. The file is decoded by your browser and transcribed by a Whisper speech model running in a Web Worker on your device. The only thing downloaded is the model itself, from Hugging Face, the first time you use it.
How big is the download, and does it happen every time?
The small model is 44 MB, the medium one 80 MB and the large one 252 MB. The size is shown before you start. Your browser keeps the model in its cache, so the next transcription starts without downloading anything.
How accurate is it?
It depends on the model and the recording. Clear speech from one person transcribes well even with the small model. Noise, music, people talking over each other, names and jargon cause mistakes. Pick the large model for difficult audio, and correct the rest in the editor: each line plays from its own start time.
How long can the recording be?
There is no minute cap or account. The limit is your device: one hour of audio takes roughly 230 MB of memory once decoded, so very long recordings can fail on phones and older computers. Splitting a long file into parts avoids that.
Which languages does it understand?
Whisper is multilingual. The language picker lists 15 common languages, and “Detect automatically” lets the model choose among all the languages it knows. Choosing the language yourself is more reliable for short clips.
Which browsers work?
Any browser with Web Workers and WebAssembly. It was verified in Chromium-based browsers (Chrome, Edge) on desktop. Other browsers and phones have not been verified yet and may be slower.