Transcribe audio
From audio or video to text with Whisper, OpenAI's speech recognition model, running entirely in your browser.
This browser can't enable the features Whisper needs (SharedArrayBuffer in an isolated page), so transcription can't start here. This mostly happens in browsers built into apps (WhatsApp, Instagram, Facebook, Telegram…): open the page in an up-to-date Safari, Chrome, Firefox or Edge. PizzaVox's other tools still work.
WebAssembly is turned off in this browser, so transcription can't start. On iPhone and iPad this happens with Lockdown Mode on (you can exclude this site in Safari's settings); otherwise update the browser or open the page in Safari, Chrome, Firefox or Edge.
This browser uses a dedicated version of the page to run Whisper. Please select your file again to continue: it still stays on your device.
Transcription happens in your browser, with no server: long audio can take several minutes, even more on a phone.
This audio is longer than 10 minutes: on a phone, transcription can take a long time and drain the battery. If you can, use a computer; otherwise pick the Fast model and keep the screen on.
Your file is processed in your browser and is never uploaded to a server.
Private transcription, no upload
Meetings, interviews, lectures, voice notes: these are often exactly the files you'd never want to upload to an unknown server. Here you don't have to. PizzaVox uses whisper.cpp, an optimised version of Whisper compiled to WebAssembly, which runs inside your browser tab.
You can pick between two models: Fast (tiny, 31 MB) to get the text quickly, and Accurate (base, 57 MB), recommended for languages other than English. The model is downloaded once and stays saved in your browser.
How to use it
Add your file
Drop an audio file (MP3, WAV, M4A, OGG…) or a video: the audio track is extracted automatically.
Pick model and language
Choose the spoken language, or let Whisper detect it. The first time, the model is downloaded and cached.
Copy or download the text
Copy the transcript or save it as .txt, or as .srt and .vtt subtitles with the timing of every sentence.
Frequently asked questions
How accurate is the transcription?
With clean audio the Accurate model is very reliable in both English and Italian. Background noise, overlapping voices and jargon can cause mistakes: you can edit the text before copying or downloading it.
How long does it take?
It depends on your computer and the model. On a recent laptop the Fast model transcribes faster than real time; the Accurate model is about two to three times slower. Keep the tab open until it's done.
Can I make subtitles for a video?
Yes: add the video directly and, when the transcription is finished, download the .srt file (or .vtt for the web). Every sentence is already timed.
Why do I need to download a model?
The model is the "brain" that recognises speech. To avoid sending your audio to any server, the model has to come to your browser instead: you download it once from our servers and then it stays cached.