Turn interviews, lectures, meetings, and voice memos into editable, timestamped text (with SRT/VTT/TXT export), entirely on your device. Nothing is ever uploaded.
Drop in an interview, a lecture recording, a meeting, or a stack of voice memos (one file or several at once) and the app reads them directly off your machine. The one thing worth stating plainly, once: none of that audio is ever sent anywhere. The speech recognition engine downloads to your browser the first time you use it and is cached from then on, so every run after that works with your internet connection switched off entirely. Turn on airplane mode after the first download if you want to check for yourself.
Files in a batch are transcribed one at a time, not because of an artificial limit but because a single file already keeps the recognition pipeline busy, and running several at once would mostly just risk running a shared device out of memory. You don't wait for a file to finish before reading it, though: as each short stretch of audio finishes processing, its words land in a live transcript on screen, selectable and copyable immediately, while the rest of the file is still being read. Cancel any file mid-queue and the ones already done stay untouched.
This is a real hardware trade, not a marketing label: one mode genuinely produces better transcripts than the other, at a real cost.
Runs the recognition model on your device's processor, no graphics hardware required. It's the default, works on essentially any laptop or phone, and is accurate for clear, single-speaker audio. Smaller one-time download.
Runs a much deeper recognition model on your graphics card instead. It needs a larger one-time download and hardware that supports it, so it only appears as an option when your browser can actually run it. There's no way to force it on unsupported hardware. In exchange, it handles accents, background noise, overlapping speech, and fast talkers noticeably better, closer to what paid transcription services produce.
A three-hour meeting recording is fully supported, and it doesn't get loaded into memory as one giant file. Underneath, a voice-detection pass slices the audio into short chunks (up to about 28 seconds each) timed to end during a natural pause whenever one falls nearby, rather than at an arbitrary fixed boundary. That matters twice over: cutting mid-word at a fixed mark loses fragments at the seam, and feeding a speech model pure silence is a known way to make it start inventing text that was never said. Pure silence gets skipped entirely rather than transcribed.
On the rare stretch of continuous speech longer than 28 seconds, the chunk still gets cut, and the next one overlaps the previous by two seconds so the words are re-stitched using their timestamps. You'd never notice the seam in the output. Progress bars track speech-seconds processed rather than file duration, so a podcast with a long musical intro reports progress honestly instead of appearing to stall on dead air.
Editing is direct: click any line, fix the text, and that line's word timings recalculate automatically so captions don't drift out of sync with the audio. There's no manual timestamp fiddling. Language auto-detects from the audio by default, or you can force one of roughly thirty languages from a dropdown if it guesses wrong.
Export covers four outputs: .srt and .vtt for captions, plain text for notes (with an optional [HH:MM:SS] prefix on each paragraph), or straight to the clipboard. Caption files follow the conventions professional captioners use by default: at most 42 characters per line, two lines per cue, and a capped reading speed so a cue never flashes by faster than someone can read it. With several files queued, export each one individually or download all of them in a single click (each file downloads separately, lightly staggered so the browser doesn't mistake them for a pop-up flood).
Format detection doesn't trust file extensions, either: the app reads each file's actual header bytes to work out whether it's WAV, OGG, FLAC, or an MP3/M4A container, so a file that got renamed or has a wrong extension still lands on the right decoder instead of silently failing at intake.
Came here for a specific kind of file? These companion guides cover the quirks of each format (the same tool, with instructions and answers tailored to what you are dropping in):