Question 1

Is my recording uploaded to be transcribed?

Accepted Answer

No. The audio is decoded by your browser and the model runs on your own machine. The only download is the model itself, from huggingface.co, the first time - that tells them which model size you chose and nothing about what you are transcribing.

Question 2

Which model should I choose?

Accepted Answer

Fast is the smallest and quickest, and fine for clear speech. Balanced is noticeably more accurate for a modest extra download, and is the default. Accurate handles accents, noise and specialist words best, but it is about 250 MB and slow without a graphics card.

Question 3

How do I make subtitles for a video?

Accepted Answer

Open the video file itself - the sound is taken from it - transcribe, and press .srt. Almost every video editor and player, and YouTube, accepts an SRT file. Each subtitle is one phrase with the times Whisper gave it, so check a few before publishing.

Question 4

Does it tell different speakers apart?

Accepted Answer

No. Whisper writes what was said, not who said it, so a conversation comes out as one stream of text with timestamps. Speaker labelling needs a separate model this page does not run.

Question 5

How accurate is it?

Accepted Answer

On clear speech, very: a test recording came back word for word. Background noise, overlapping voices, strong accents and unusual names are where it slips, and the larger models slip less. It can also occasionally repeat a phrase or invent one in a long silence, so treat it as a first draft.

Free audio to text transcription