A transcriber that never sees your audio.
Speech to text and subtitles, inside your browser. Why, how, who runs it, how the numbers are made, and where it falls short.
Why it exists
Most transcription services upload your recording to their server, run a model there, and send the text back. You trust them with a meeting, an interview or a therapy session, and read a policy to learn how long they keep it. They also pay for every minute, which is why free tiers stop at thirty minutes a month.
Speech recognition no longer needs a server. The open-source Whisper models run on an ordinary graphics card, and browsers now expose that card through WebGPU. So the model runs in the tab and the upload step disappears, along with the policy, the retention and the cap. That is the whole idea.
What we care about
- Nothing leaves the browser. There is no upload path in the code and no server to receive one.
- Free with no meter. No account, no time limit, no paid tier. Local processing costs nothing per file.
- No friction. Drop a file, watch the text appear, download it.
- Exports that work. Text for notes, SRT for video platforms, VTT for web players, with sensible cue lengths.
- Honest numbers. Every speed or accuracy claim comes from a run we did, on named hardware, on audio you can hear.
How it works
The browser reads your file in slices and WebCodecs decodes it a few seconds at a time, mixed to mono at 16 kHz. The model gets windows of about a minute, each cut at a quiet moment, and the text for each window appears as soon as it is ready. Only a minute of audio is held at a time, so a three-hour file uses about as much memory as a three-minute one.
The default model is the base English size of OpenAI's open-source Whisper, run in the browser with transformers.js and ONNX Runtime Web: about 200 MB, downloaded once. With WebGPU it runs on the graphics card; without, on the processor, more slowly. The browser support page checks yours.
| What | Where it goes |
|---|---|
| Your audio or video file | Nowhere. It stays in the browser tab and is released when you close it. |
| The transcript | Nowhere, until you copy it or download it yourself. |
| The model weights | Downloaded to your browser cache once, then reused. |
| Page analytics and ads | Ordinary web requests to Google, subject to your consent where consent is required. See the privacy policy. |
Who runs it
FreeTranscribe is built and run by Alain, a software developer in Europe, as a side project. No company, no investors, no affiliation with OpenAI or any transcription service. The editorial policy says how the articles are written and checked, including where AI assists.
How claims are tested
We quote no speed or accuracy figure we have not measured. The examples page shows the recordings, the tool's output from a real browser run, the reference, the word error rate and the machine. Missing measurements say so. The one number we can give today: on a 2021 laptop running the model on its integrated Intel graphics in Chrome, the 71-second clean-speech sample took 25 seconds with the model already cached, with no word errors against the reference. That clip is synthetic speech and flatters the model; real recordings will do worse.
Limitations
- Browsers. Desktop Chrome and Edge are what we test on. An iPhone 17 works in phone mode, the small model on the processor. Desktop Firefox and Safari ship the needed APIs but are unverified, and Android phones are untested. Browsers without WebGPU get the slower CPU mode.
- Speed depends on your hardware. A discrete graphics card is comfortably faster than the recording; a laptop with integrated graphics takes roughly as long as the recording; CPU mode is slower again. We have measured one machine so far.
- Accuracy. The base model gets clean English speech right and stumbles on names, technical terms, heavy accents, crosstalk and noisy rooms. It can invent text over silence, which we filter but cannot fully prevent. Read every transcript before relying on it.
- No speaker labels. The transcript does not say who is speaking. Two-person interviews come out as one stream of text.
- Other languages and the larger model have one test run each. The picker also offers a multilingual base model, tried once on a Dutch synthetic clip, and the larger turbo model, which on our laptop took about as long as the recording and fixed the base model's one error on noisy audio. Both rows are on the benchmarks page; neither has been tested on real recordings yet.
- Microphone mode is new. Words appear a few seconds after you say them, in windows cut at pauses, so a sentence can occasionally be split across two lines. It has been tested with a recorded clip fed through the browser's microphone input, not yet with many real microphones.
- Very long files without WebCodecs. On the Web Audio fallback the whole file is decoded at once, so recordings over roughly an hour may fail there.
How it is paid for
A few ads cover the domain and hosting. There is no paid tier and no plan for one. Ads stay away from the drop zone, the transcript and the download buttons, and are not shown when the tool cannot run in your browser.
Contact
[email protected], or the contact page.