Transcribe a video to text, in your browser
Open a video or an audio file and its speech becomes text: copy the transcript, or save it as a text file or as SRT or VTT captions timed to the speech. It hears English, Spanish, Italian, French, Portuguese and German. A speech model runs in your browser, so the file is never uploaded.
Transcribe your video or audio
Your file stays on this device. Any video or audio file this browser plays opens here, up to 30 minutes long. The first time, the speech model downloads, 94 MB, or 125 MB to run on the graphics card. Tick the box to keep it in this browser for next time.
How to transcribe a video or audio file
- Choose the language spoken in the recording: English, Spanish, Italian, French, Portuguese or German.
- Open the file: an MP4, MOV or WebM video, or an MP3, M4A or WAV recording. The sound is read with your browser’s own decoder.
- The speech model downloads first: once if you keep it in this browser, or again each time you open the page if you don’t. Then the bar follows the transcription, and Cancel stops it.
- Press Copy text, or Save .txt for the transcript, or Save .srt or Save .vtt for captions. Each file is named after yours.
What runs in your browser
| Part | What it is |
|---|---|
| Speech model | Whisper base by OpenAI, the multilingual model, 74 million parameters |
| Languages | English, Spanish, Italian, French, Portuguese and German. You choose, and the model is told rather than left to guess |
| Conversion | Xenova/whisper-base, the model in ONNX format, 8-bit, with a half-precision encoder for the graphics card |
| Runtime | Transformers.js on ONNX Runtime Web, on WebAssembly with one thread, and on WebGPU too where the browser offers it |
| Download | 94 MB, or 125 MB with WebGPU, from Yaptake’s model host: once if you keep it, otherwise once per visit |
| Speed | A one-minute clip in 15 to 20 seconds on an M5 Max Mac without WebGPU, and 6 to 10 with it |
Licenses: Whisper is MIT, the ONNX conversion and Transformers.js are Apache 2.0, and ONNX Runtime Web is MIT. Their notices are in speech-licenses.txt.
How it works
The tool reads the sound of your file at 16 kHz in mono, the way the model hears it, and hands it to the model in a worker of its own, so the page keeps responding. The model is told the language you chose and writes down what it hears in it. It reads 30 seconds at a time, writing the words and timing each stretch of speech. The end of those 30 seconds may cut the last stretch in two, so the next 30 start where it began: each sentence is read whole, and once.
The transcript starts a new paragraph after a pause of 2 seconds or more. The captions are those same stretches of speech, usually a sentence each, at the times the model gave them. A stretch that only names a sound, such as [Music], is left out of both, and so is anything the model writes where the recording is silent.
Your file stays on your device
Your video or audio file is never uploaded: the speech model runs in this browser, and the transcript is made there. The model’s files come from Yaptake’s own model host on Cloudflare, and each is checked against the fingerprint built into the tool before it runs. Your browser keeps them only if you tick “Keep the speech model in this browser”. In the European Union, the European Economic Area, the United Kingdom and Switzerland the box starts unticked, and elsewhere ticked; untick it to remove the kept model.
The site counts that the tool was used, in broad bands: the recording’s length, how many words were found, whether the model ran on WebGPU or WebAssembly, whether it was downloaded or already kept, and whether you chose to keep it. It never sends the file, its name or any of its words, as the privacy policy describes.
Limits
- A small model. It can mishear names, jargon and quiet or overlapping voices, so read the text before you publish it. In our tests it made more mistakes in German, French, Portuguese and Italian than in English and Spanish.
- It writes in the language you choose. Told the wrong one, it writes the wrong words, a loose translation or nonsense, so choose the language spoken: changing it reads the file again.
- Recordings up to 30 minutes long and 512 MB.
- Yaptake transcribes on your Mac with a larger speech model, Parakeet, which it downloads once, in the same six languages.
Questions
How do I convert an MP4 to text?
Open the MP4 in the tool above and wait for the transcript, then press Copy text or Save .txt. The video plays no part: only its sound is read.
How do I transcribe an audio file to text?
The same way as a video: open the MP3, M4A, WAV or other recording in the tool above, then press Copy text or Save .txt. Any sound file your browser plays opens, up to 30 minutes long.
How do I make an SRT file from a video?
Open the video, wait for the transcript, and press Save .srt, or Save .vtt for WebVTT. The captions are timed to the speech. To move them all earlier or later, open the file in the subtitle converter.
Which languages does it transcribe?
English, Spanish, Italian, French, Portuguese and German, the six Yaptake transcribes. Choose the one spoken before you open the file; change it after, and the file is read again. The model doesn’t guess: it writes in the language it is told.
Is my video or audio uploaded?
No. The speech model runs in your browser, and the file never leaves your device.
Why does the first transcription take longer?
The speech model downloads first, 94 MB, or 125 MB to run on the graphics card. If you keep it in this browser, after that the tool starts in about a second.
Why did it find no speech?
The recording may be silent, or hold music or other sounds rather than speech. What the model writes over a silent stretch is left out.
Which browsers does it work in?
It needs a browser with WebAssembly, such as a recent Chrome, Edge, Firefox or Safari. Where the browser offers WebGPU, the model runs on the graphics card, which was twice as fast on the Mac we measured.
Try Yaptake
Yaptake transcribes on your Mac with a larger speech model, in English, Spanish, Italian, French, Portuguese or German, and edits the video itself by its transcript: delete a sentence and its picture and sound go with it. It makes captions from the speech and draws them into the video, with nothing uploaded.
Your first 3 successful exports on each Mac are free, with no watermark, account or credit card required. After that, exporting is $9.99 a month or $79 a year. Editing is always free.
Download for MacEditing by textSubtitle converterAll free tools
