iLoveAudioVideo

Whisper large-v3-turbo online, on your own computer

Run OpenAI's Whisper large-v3-turbo model without an API key, a server or an install. It runs in this browser tab on your own graphics chip, so your recording is never uploaded and there is no limit on length or number of files.

  • Never uploaded
  • No sign-up
  • No daily limits
  • Free

Quick answer: To use Whisper large-v3-turbo online, drop an audio or video file into the tool at the top of this page and choose Best quality: Best is Whisper large-v3-turbo. The first time, the browser downloads the model (about 566 MB) and keeps it, then it transcribes on your own graphics chip. You get text, SRT or VTT. Nothing is uploaded.

Updated

How it works

  1. Drop a recordingMP3, M4A, WAV, MP4, MOV and most other audio and video files. It stays on your computer.
  2. Choose Best qualityBest is Whisper large-v3-turbo. It appears when your browser can use your graphics chip, as Chrome and Edge do on most recent computers.
  3. Read, fix and downloadClick any word to correct it, then save TXT, SRT or VTT.

Which Whisper model each quality uses

QualityWhisper modelFirst download
Fasttinyabout 44 MB
Balancedbaseabout 80 MB
Highsmallabout 252 MB
Bestlarge-v3-turboabout 566 MB, graphics chip only

Each model downloads once and is kept by the browser, so the next file starts straight away, even offline. Auto picks the most accurate one your computer runs at a good speed; pick Best yourself to force large-v3-turbo.

Large-v3-turbo vs large-v3

Large-v3-turbo is OpenAI's faster version of large-v3: the same encoder, which does most of the listening, with a much smaller decoder that writes the words. OpenAI reports that it is several times faster with accuracy close to large-v3 for most languages. That speed is what makes it practical inside a browser tab. The full large-v3 is not offered here: its decoder is eight times bigger, which makes it much slower inside a browser tab.

How fast it is, and when it is not offered

On an Apple M1 test machine, Best ran at about 1.9 times real time: an hour of audio in roughly half an hour. Speed depends on your graphics chip.

Best needs a browser that can use the graphics chip. If yours cannot, or the computer has very little memory, Best is not shown, because running this model on the processor alone needs a 1.1 GB download and froze an 8 GB laptop in our tests. High quality (Whisper small) still works there.

Measured, not promised

~4%of words wrong on clear English with Balanced quality (about 3% with High)
9–28 minfor a one-hour recording on the computers we tested
Any lengthhour-long recordings finished on every computer we tried

Results from our benchmark on clear, read English speech. Noisy audio, heavy accents and crosstalk give more errors. See the full benchmark.

Why your file never leaves your computer

Your recording is transcribed by your own computer, inside this tab, and never uploaded. That is why there is no daily cap and no length limit. How it works.

Questions

Is this the real OpenAI Whisper model?

Yes. It is OpenAI's open-source whisper-large-v3-turbo, in a compressed web version (smaller numbers inside, so it fits a browser), running in your browser. No cloud service or API is involved, so there is no per-minute price.

Do I need an API key or an account?

No. There is nothing to sign up for. The model files download straight to your browser the first time you choose Best.

Does it label speakers?

Not yet. You get the words with timestamps. Telling speakers apart is on our list.

Can I use it offline?

Yes, once the model has downloaded. Keep the page open and transcribe without an internet connection.

Which quality should I choose?

Leave it on Auto. It picks the most accurate quality your computer runs well, and switches to a lighter one if your computer slows down. Best needs a fast computer.

Is my audio uploaded anywhere?

No. The file is read and transcribed inside your browser tab. The speech engine is downloaded once.