Turn speech in a video or audio file into editable subtitles, recognised by an AI model that runs on your own device. Nothing is uploaded, and there’s no account or watermark.
Drop a video (MP4, MOV, WebM, MKV) or an audio file (MP3, M4A, WAV, OGG) on the page, or tap to choose one. The page shows its length and size.
Leave “Spoken language” on “Detect automatically” or pick the language yourself, choose whether the “Subtitles in” the spoken language or “English (translated)”, and pick an “Accuracy” level.
Press “Generate subtitles”. The first run shows “Downloading the speech model…”, then “Transcribing…” with a percentage and the time left, and lines appear as they are recognised.
Correct any words or times in the subtitle list, check them in the “Preview”, then use “Download SRT”, “Download VTT” or “Download TXT”.
filetidy uses Whisper, the open speech-recognition model published by OpenAI, and runs it inside your browser with transformers.js. On a browser with WebGPU the model runs on the graphics card; elsewhere it runs on the processor through WebAssembly. Your file is never sent to a server: the audio is read, recognised and turned into subtitles on your device.
Before recognition, the sound track is pulled out of the video in the browser and resampled to 16 kHz mono, the format Whisper was trained on. Whisper then predicts words together with timestamps, and filetidy groups them into subtitles that break at sentence punctuation and natural pauses, last roughly 1 to 7 seconds, and fit the line length you set.
Whisper knows around 99 languages, but its quality is uneven. Widely spoken languages recorded with a clear voice come out close to perfect; rare languages, strong accents, background music and people talking over each other cause mistakes. Names and numbers are the usual weak spots, so read the result before you publish it.
Like any model of its kind, Whisper can occasionally invent words during silence or music. filetidy skips the silent stretches before recognition to keep that to a minimum.
Windows
Live captions only display text on screen; nothing is saved as a subtitle file, and the list of supported languages depends on your Windows version.
Mac
Live Captions needs a Mac with Apple silicon, is focused on English and isn't offered in every country or region. It produces no file you can edit or upload.
Android
Menu names vary by brand, and which languages are available depends on the phone. Captions stay on screen and are not exported.
The “Accuracy” setting picks the size of the model. “Fast” is Whisper tiny, a download of about 44 MB (about 106 MB with WebGPU), and is the default on phones. “Balanced” is Whisper base, about 80 MB (about 170 MB with WebGPU), and the default on computers. “Best” is Whisper small, about 400 MB, and only appears when your browser supports WebGPU because it would be too slow on the processor alone.
Each step up catches more words correctly, especially in languages other than English and in noisy recordings, but takes longer. On a recent laptop with WebGPU, several minutes of video are processed in well under a minute; on the processor, Balanced runs at about the speed of playback or a little faster. Phones take longer.
Start with the default. If the result has many errors, run it again with a larger model, and set the spoken language instead of relying on detection. Those two changes fix most problems.
No. Speech recognition happens in the browser tab, so the video or audio stays on your device. The only things downloaded are the speech model and the engine that runs it, and your browser keeps them in its cache so later runs start straight away.
Whisper covers around 99 languages, from English, Spanish, French and Arabic to many smaller ones. Results are best for languages with lots of training data and clean recordings; for less common languages, use “Best” if it is offered and expect to correct more words.
Yes, into English only. Set “Subtitles in” to “English (translated)” and Whisper writes English subtitles directly from speech in another language. Translating into any other language is not something the model can do.
The free version processes up to 10 minutes of audio per file. If your file is longer, the page tells you so and subtitles the first 10 minutes. A Pro version for files of several hours is in preparation; its waitlist will open on the pricing page.
The on-device version is free, needs no account and adds nothing to your files. The subtitles you download are plain text files that belong to you.
With a clear voice in a common language, most sentences are right on the first pass. Accuracy drops with music under the speech, several speakers at once, heavy accents, product names and figures. Treat the output as a first draft and proofread it in the editor.
Pro is coming soon. It will run Whisper large-v3 turbo on filetidy's servers, which is much faster, handles accents and dialects better and accepts files that are hours long. Only compressed audio would be sent, never the video, and it would not be stored.
Last updated October 28, 2026
Choosing the language is faster and more accurate than detection, especially for short clips.
Speech recognition makes mistakes, especially with names, numbers, music and several people talking at once. Dialects such as Moroccan Darija are recognised less reliably than Modern Standard Arabic. Read the subtitles before you publish them.
Need long videos or the best accuracy?
AI subtitles Pro will use Whisper large-v3 on our servers: about an hour of video in a few minutes, better accuracy on accents and dialects, files up to 3 hours. Only compressed audio is sent, never the video.
See Pro and join the waitlist