About this tool
Turn a recording into text without sending it anywhere. OpenAI's Whisper speech recognition runs inside your browser through ONNX Runtime: drop a voice memo, lecture, interview, podcast, or meeting recording and get the transcript as plain text, with SRT and VTT subtitle files that carry timestamps. Batch mode transcribes a folder of recordings one after another.
Choose the model for the job. Tiny is small and quick for clear English; Base understands more than ninety languages and can translate any of them into English; Small gives the best accuracy at the cost of time. The model downloads once and is cached, so later transcriptions start immediately and work offline.
How to use it
- Drop your MP3, WAV, M4A or OGG files into the box above, or click to browse. Nothing is uploaded.
- Adjust the options if you want; the defaults suit most files.
- Click the button and wait a moment.
- Download the results individually or all at once as a ZIP.
Questions
How accurate is it?
Tiny is fine for clear English speech and gets names and jargon wrong; Base handles accents and other languages; Small is close to what paid services offer but takes several times longer. All run on your device, so a faster computer helps.
How long does it take?
In a browser with WebGPU (Chrome, Edge, and recent Safari on most computers) the model runs on the graphics card and an hour of audio takes a few minutes with Base. Without it, the CPU path transcribes Tiny about as fast as the audio plays, Base two to three times slower, and Small much slower: a one-hour recording with Tiny is a coffee break, with Small an afternoon.
Is the audio uploaded?
No. The model file is downloaded to your browser; the audio is decoded and transcribed on your device. The network panel shows only the model download.
Can I get subtitles?
Yes. Every transcript comes with SRT and VTT files with timestamps, ready for video players, YouTube, or the subtitle tools here.
Is it free, are files uploaded, which browsers work: the general questions are answered once for the whole site.