Live transcription
Transcribe has a second mode: instead of files, it listens to a microphone and writes down what is said, a phrase at a time. Everything runs on your computer with Whisper.
Switching to the microphone
At the top of Transcribe, the What to transcribe switch has two options: Files and Microphone. Choose Microphone, then set:
- Language spoken — the language you will speak.
- Speech model — smaller models keep up with speech better. The screen recommends one and warns when your choice is too slow for live use.
- Microphone — Default microphone, or a specific device. Device names appear once the microphone has been allowed.
- Vocabulary — names and terms to spell correctly, as for files.
Live transcription uses Whisper only. The first time, your system asks permission to use the microphone: macOS shows its usual prompt, and on Windows the app asks through the system.
How it works
Click Start listening. The app captures your voice and cuts it at pauses — after about 0.7 seconds of quiet, or after 12 seconds of continuous speech — and recognises each phrase with its own Whisper run. The text appears as each phrase is finished, with a level meter and a running clock while you speak.
It is phrase by phrase, not word by word, and each phrase is recognised without the one before it. Because the model loads for every phrase, the recommendation never includes the Large model, and not Medium on a processor without a graphics card: a model that keeps up beats one that is slightly better.
Stop ends the session; the app finishes the phrase being recognised first. Continue listening starts a new session whose times carry on from where the last one ended instead of restarting at zero. The transcript survives leaving the screen, and an open microphone is treated as running work, so the app asks before quitting.
Copying and exporting
Copy text puts the transcript on the clipboard. Export… saves it as a subtitle file or plain text — the extension you choose in the save dialog decides the format: SRT, WebVTT, ASS or TXT. Times come from the amount of audio captured, not the clock, so a slow model changes when text appears on screen but never what the exported file says. Clear empties the transcript.
Sharing the machine
Live phrases share the processor or graphics card with any files being transcribed. A phrase waits up to 20 seconds for its turn; if a file is still holding the graphics card after that, the session ends with a message. For live use, finish or stop file transcription first, or pick a smaller model.