Transcribe recordings into editable text with a Whisper model running on your device: nothing uploads, timed lines edit in the page, and TXT, SRT and VTT downloads rebuild from your edits.
Your files are processed in your browser, on your own device. Nothing is uploaded to a server.
Speech to text turns a recording into words you can edit, search and publish: voice memos, interviews, lectures, meetings. This tool transcribes on your device with a Whisper model that downloads to your browser, which makes the privacy claim simple and total: the model comes to the recording, and the recording never leaves the machine.
Set that against the shelf it sits on. Server transcription services take your file onto their computers, keep transcripts and request data in operational logs by their own privacy lines, and charge by the minute past a free taste. Here there is no upload to delete later, because there is no upload at all. The cost is stated just as plainly below: the model is a real download, and on a laptop without graphics acceleration, long files take real time.
1. Drop the recording into the box: MP3, WAV, M4A, OGG, Opus, FLAC and friends, up to 200 MB.
2. Pick the model (Tiny for speed and long files, Base for better accuracy on shorter, clear speech) and the spoken language, or let it detect.
3. Press Transcribe, then edit the timed lines in the page and download TXT, SRT or VTT rebuilt from your edited text, or copy it whole.
Tiny downloads about 43 MB and Base about 80 MB, once each, then the browser caches them, and both figures are printed before anything downloads. Tiny runs several times faster and suits long recordings. Base is the better ear for clear speech under about 30 minutes. On a CPU-only laptop Base can take about as long as the recording itself, and where your browser and device offer WebGPU, both shorten considerably, device-dependent and never promised as a number here.
Accuracy, in the register this page commits to: very usable on clear speech with a decent microphone. Weaker on overlapping voices, background noise, far-away speakers and accents the model knows less well. No percentage is printed, because a single accuracy figure for all recordings is a marketing invention. There are no speaker labels: these models do not separate speakers, and transcripts arrive as one timed stream. A translate option renders English text from other languages, done by the same model, labelled as such.
This page owns audio files and never accepts video. video to text owns video files and extracts the soundtrack first. The engines share the model code and differ only in intake, and each page sends the other file type onward in one line, as this one just did. Record something fresh with the voice recorder and it drops straight in here. The full family is in the audio tools collection.
Every transcript, from any service, is a draft until a human who knows the names reads it. Names, jargon and numbers are where models guess, and the timed-line editor exists for exactly that pass: fix the wording in place, and the downloads rebuild from your edited text when you press them, so the file you keep is the text you approved, never the raw first pass.
Very usable on clear speech with a decent microphone, weaker with overlap, noise, distance and under-represented accents. No percentage is promised, because accuracy belongs to your recording, not to a landing page.
Yes, that is the architecture: a Whisper model downloads to your browser (about 43 MB for Tiny, 80 for Base, cached afterwards) and the recording is processed on your device. Nothing is sent anywhere.
Files up to 200 MB. As working guidance, Tiny is the choice for long recordings and Base for shorter clear ones: on a CPU-only laptop Base can take about as long as the recording, while Tiny runs several times faster.
No. These models produce one timed stream of text without speaker labels. Services printing Speaker 1 and Speaker 2 run additional models this tool does not ship, and the page would rather say so.
Yes. The transcript downloads as SRT and VTT with timings, rebuilt from your edited lines, alongside plain TXT and a copy button for the text alone.
Use Video to Text in the video tools: it accepts video files and pulls the soundtrack out first, then runs the same models. This page accepts audio files only.
Same rules as this one: free, in your browser, with nothing uploaded.
MP3, WAV, M4A and OGG, swapped in seconds.
Cut a ringtone or a clip out of any song.
Heavy WAV files made small enough to share.
Voice memos and iTunes files as plain MP3s.
An MP3 opened back up into a full WAV file.
OGG files, often from games and apps, as MP3s.
Big FLAC songs made small enough for a phone.
An MP3 saved as OGG for games, apps and the web.