Audio to text

• The speech model is OpenAI’s Whisper, trained on 680,000 hours of multilingual speech. The page runs Whisper base (74 million parameters) or tiny (39 million) in the browser. Source
• Voice Memos on a Mac records Compressed (smaller files) or Lossless (larger files), set in Voice Memos › Settings › Audio Quality. Compressed recordings are AAC in an M4A file. Source
• Microsoft 365 subscribers can transcribe at most 300 minutes of uploaded audio a month with Transcribe in Word for the web. Source
• English speakers in the United States average about 150 words a minute, so an hour of steady talk is roughly 9,000 words of transcript. Source

Choose the MP3 or drop it on the box above. The page reads its length and checks that it can decode the sound before anything else is downloaded, so a damaged or unusual file is caught in a second rather than after a long wait. Leave the language on Detect automatically, or pick it if the first half-minute is music or a different language from the rest.

Press Transcribe. The first time, the speech model downloads and your browser keeps it, so the next file starts at once. Lines appear about every half-minute of sound, with the time left underneath. When it finishes, play the recording from any line, fix names and terms, and press Download .txt. The same steps work for M4A, WAV, OGG, FLAC and the sound of a video file.

The work happens on your device, so speed depends on it. Before you start, the page says how long it expects the whole recording to take, in minutes, from the kind of device you are on; once it has run once, it uses the speed it measured on yours. A recording that would take more than 20 minutes is offered online transcription as well, which you can take or leave.

These figures come from our own tests and from visitors’ runs of the same engine, for one hour of sound. Your computer may be faster or slower, and the model download is extra the first time.

You may already have a transcriber. These are fair alternatives, each with its own trade-off.

• Word for the web (Microsoft 365): open a document, go to Home › Dictate › Transcribe, select Upload audio and pick the file. The file is uploaded to Microsoft, speakers are labelled Speaker 1, Speaker 2 and so on, and subscribers get 300 minutes a month.
• Notes on iPhone: on an iPhone 12 or later with the device language set to one of ten supported languages, Notes can record audio and show the transcript as you record. It suits the next lecture or meeting you record rather than a file you already have.
• Whisper on your own computer: the same model family, run from a terminal with Python. Install it with pip install -U openai-whisper, then run whisper recording.mp3 --model turbo. It also needs ffmpeg installed. It is free and offline, and the larger models are more accurate than those that fit in a browser, but it needs a capable computer and some setup.
• Paid transcription services: services such as Otter, Rev and Descript upload the file, label speakers and often add summaries. They charge by the minute or by the month, some with a small free allowance.