Create subtitles for a video without uploading it
Most people watch social videos without sound, and recorded meetings are easier to skim with a transcript. Automatic subtitles used to mean uploading the recording to an online service. Today a speech recognition model can run directly in your browser.
Updated:
How it works
The subtitle generator runs the Whisper tiny model from OpenAI on your device. It detects the spoken language automatically, supports 99 languages, and can translate speech into English subtitles if you want.
The model is downloaded once, about 41 MB, and then stored by your browser. After that, subtitles can be created offline, and neither the video nor the audio leaves your device.
Step by step
Both video and audio files work, for example a phone recording, a podcast episode or an exported meeting.
- Open the subtitle generator and select the video or audio file.
- Keep automatic language detection, or pick the language if the recording starts with music or silence.
- Start the recognition. On a current laptop it usually takes less time than the recording lasts; phones are slower.
- Read through the lines and correct names, technical terms and timings in the editor.
- Download SRT, VTT or plain text, or create a new video with the subtitles burnt in.
SRT, VTT or burnt in?
SRT is the most widely supported subtitle file. Video players, editing programs and most upload forms for video platforms accept it. VTT is the format for web video players. Plain text is useful for minutes, notes or searching what was said.
Separate subtitle files can be switched on and off and edited later. Burnt-in subtitles are part of the picture, which is what you need for platforms that do not accept subtitle files. UseTool writes such a video as a new WebM file and leaves the original untouched.
Getting better results
The tiny model is small enough to run in a browser, and that has limits. Clear speech close to the microphone works well. Background music, several people talking at once, strong accents and heavy reverb lead to more mistakes.
Names and technical terms are the most common errors, so it is worth reading the text once before publishing. Setting the language manually helps when a recording mixes languages or starts with a long intro.
Recordings of up to an hour are supported. On phones, very long recordings can run out of memory, so shorter parts are more reliable there.
Why local processing matters here
Recordings of meetings, interviews, lessons or therapy sessions often contain personal information about people who never agreed to share it with a transcription service. Processing on your own device avoids that question entirely, and there are no minutes counted against a quota.