Video to text. On your device.

No. Audio is decoded on your device, then a Whisper model runs in a browser worker. Model files are downloaded from Hugging Face, which receives normal download requests, but your recording and transcript are never sent there. The inference engine is served by this website.

It is free and needs no account or API key. Your device does the work. Standard model weights are about 41 MB for Faster, 77 MB for Balanced and 249 MB for Best. Best also offers an optional GPU acceleration configuration with about 322 MB of weights. First use adds language files, the inference engine and, by default, a 2.3 MB speech detector. Downloads begin when you transcribe, and cached assets can be reused while browser storage retains them.

Best uses a larger Whisper model, so it needs more processing time and memory. CPU processing uses up to four threads when supported. For Best, Processing options also offers GPU acceleration (~322 MB), requiring WebGPU with 16-bit floating-point support. Compatibility is checked before downloading accelerated weights. If unavailable, standard CPU processing is used; a later GPU error may require another 249 MB download for CPU fallback. Leave Detailed subtitle timing off for faster processing. Speed depends on your device and recording, so compare a short sample first.

MP4, MOV, MKV, WebM and common audio containers are supported when their audio codec can be decoded by your browser. MP4 with AAC audio, MP3 and WAV are good starting points. The current limits are 500 MB and 30 minutes per file. Recent desktop Chrome or Edge is recommended; large models can exceed memory on mobile devices.

Whisper generates text incrementally. We show those words as they arrive and add timestamped segments as they become available. The panel follows new subtitles automatically. Scroll up to pause following, then scroll to the bottom or select Back to latest to resume. Each audio section is finalized before moving to the next, so draft words and times can change. Pause takes effect after the current section; Stop & keep preserves available subtitle segments as a partial transcript.

Yes. Once transcription completes or you stop and keep a partial result, edit the words and the start and end times in seconds. Download SRT for editing tools, VTT for web players, or TXT for plain text. AI text and timing need review, especially with noise, music, accents or multiple speakers.

This tool transcribes speech in its original language. English is selected by default and uses a small English-only model. Choose another language or Auto detect to load the multilingual version instead. It does not translate, label speakers or burn subtitles into the video. For consistent timing changes after export, use Subtitle Time Shifter.