Audio to text questions
Here is how local browser transcription works and what to expect from the first version.
Does my audio leave my device?
No. The local workflow decodes the file and runs Whisper in your browser. The file is not sent to this website.
Do I need to download anything?
You do not install software. The first transcription downloads a model into the browser cache. Later runs can reuse that cached model.
Which files are supported?
MP3, WAV, M4A, and MP4 files up to 200 MB or 20 minutes in the initial workflow. Browsers with limited native codec support use a local ffmpeg.wasm fallback.
Does it recognize multiple languages?
The multilingual Whisper model can process many languages. Auto detection works best when a recording has one primary language; mixed-language recordings may need a manual language choice.
Does it identify speakers automatically?
Not in this first local build. You can assign Speaker 1 through Speaker 4 to each transcript segment during review. Overlapping speech and background noise still need manual checking.
Does it show model confidence for every word?
The current browser Whisper build returns segment text and timing, but not token probabilities. The editor therefore does not invent confidence scores; it uses glossary checks to focus manual review.