Transcribe guide
How to transcribe audio to text for free, without uploading it
Open the recording in a speech-to-text tool, pick its language, and correct the text while it plays back. Transcribe runs OpenAI’s open-source Whisper model in your browser, so the recording never leaves your device, then saves plain text, Markdown, or SRT and VTT subtitles.
Why transcribe on your own device
Most transcription services upload your recording to their servers. That’s a real problem for the recordings people most need transcribed: job and research interviews, therapy and medical notes, legal calls, team meetings, a family member’s stories.
Transcribe downloads the speech model to your browser once (about 77 MB) and runs it there. After that it works offline, and the only copy of your recording stays on your device.
Step by step
- Open the file: MP3, M4A, WAV, MP4, MOV, WebM and more, or press Record.
- Pick the language, or leave Auto. Choosing it helps with short recordings.
- Transcribe. Long recordings are cut at quiet moments and done in 30-second parts, so the text appears as it goes and you can stop at any point.
- Correct it while it plays. Tap a line’s time to hear it; type to fix a word. Find and replace fixes a name everywhere.
- Save it as text, text with times, Markdown, or SRT or VTT subtitles.
Getting the best accuracy
- Record close to the speaker. A phone on the table between two people beats a laptop across the room.
- Cut the noise. Music, a TV, wind and crosstalk cause most errors.
- Set the language for recordings under a minute, where detection has little to go on.
- Fix names once. The model doesn’t know your colleagues’ names. Replace each one across the transcript, or keep a word list with Pro so it’s fixed in every transcript.
- Try Accurate (Pro) for accents, noisy rooms and specialist words. It’s Whisper’s larger model: slower, but it makes fewer mistakes.
What to expect
On 31 seconds of clear recorded speech, the Fast model got about 88 words in every 100 exactly right and the Accurate model about 90, with most misses being numbers written as words or small words like “a”. Real recordings with background noise do worse, so always read the transcript through.
Subtitles for YouTube and video editors
Export SRT for YouTube, Vimeo, Premiere Pro, Final Cut Pro and DaVinci Resolve, or VTT for the web’s <track> element. Lines are split to two rows of about 42 characters and at most six seconds, so they read comfortably. With Pro, Transcribe can also burn captions into the video itself for Reels, TikTok and Shorts, with a word-by-word highlight.
What it can’t do yet
- It doesn’t tell speakers apart: add names to the lines yourself.
- Very long files use a lot of memory. Up to about two hours works well on a computer; split longer ones.
- Safari runs the model on one processor core, so it’s slower there than in Chrome or Edge.
Questions
Is it really free?
Yes: transcribing any length with the Fast model, editing and every export are free. Pro adds the Accurate model, burned-in captions, translation into English, a word list and unlimited saved transcripts.
Is my recording uploaded?
No. The speech model runs in your browser. Only the model itself is downloaded, once.
Which languages does it support?
Whisper knows about 100 languages, including English, Spanish, French, German, Portuguese, Japanese, Chinese, Hindi and Arabic. Accuracy is highest for widely spoken languages.
Can it translate?
With Pro, it can turn speech in any of its languages into English text.
How long does it take?
On a recent laptop the Fast model works at about two to three times real time, so a ten-minute recording takes three to five minutes.