Automatic subtitles
Choose a video (or an audio file) and press Generate: an AI speech-recognition model listens on your device and writes each sentence with its start and end time. Download the result as an SRT file, the format YouTube, Premiere, DaVinci Resolve, CapCut and VLC take, or as WebVTT for web players.
The language is detected automatically, or you can choose it. Each subtitle follows the natural pauses in speech.
Before you publish
Read through the subtitles once: names, technical terms and noisy passages are where speech recognition slips. An SRT file is plain text, so any text editor can fix a word; video editors also let you adjust timings.
Private by design
The video isn't uploaded: the sound is decoded and transcribed in your browser. The first time, the AI model (about 136 MB) is downloaded and kept by your browser. Videos up to 90 minutes long.
How this works
The video's sound is converted to 16,000 samples per second and processed in 30-second windows by Whisper base (OpenAI, MIT licence), which predicts the text together with timestamps every time speech starts or pauses. Each pair of timestamps becomes one subtitle cue, written in SRT (HH:MM:SS,mmm) or WebVTT (HH:MM:SS.mmm) timing format.
Frequently asked
Is my recording uploaded?
No. Speech recognition runs in your browser on this device; the file never leaves it.
Which languages work?
Around 100, including English, Portuguese, Spanish, French and German. It detects the language by itself, or you can choose it.
How long can the recording be?
Up to 90 minutes per file. A laptop transcribes roughly six minutes of audio per minute; the first time also downloads the AI model (about 136 MB).
How accurate is it?
Very good for clear speech by one person at a time. Background noise, music, strong accents and people talking over each other lower the accuracy, so read the text before you use it.
Can I burn the subtitles into the video?
Not on this page: it creates the subtitle file. Upload the SRT next to your video on YouTube, or import it into your video editor.
Can it go faster?
Yes, with a free account: choose Fast, in the cloud. Whisper large-v3-turbo transcribes an hour of audio in a few minutes for about 0.5 credit a minute (free accounts get 50 credits a month). The audio is sent in two-minute pieces and not stored. On this device stays free and private.
Sources
- OpenAI Whisper — robust speech recognition via large-scale weak supervision (MIT)
- Radford et al. (2022) — Robust Speech Recognition via Large-Scale Weak Supervision
- ONNX Runtime Web — run ONNX models in the browser (WebGPU and WebAssembly)
- Mediabunny documentation — Converting media files
- W3C — WebVTT: The Web Video Text Tracks Format
- Wikipedia — SubRip
Reviewed Oct 19, 2026 · test vectors fixtures/tools/video/subtitle-generator/