What Whisper is
Whisper is a speech recognition model from OpenAI, described on its own page as general-purpose: it transcribes speech in many languages, translates it into English and tells which language is being spoken. The code and the model weights are released under the MIT License, so anyone can download and run it.
It comes in six sizes, from tiny to large, trading speed for accuracy. OpenAI lists the memory each needs on a graphics card, from about 1 GB for the smallest to about 10 GB for the largest, and says its accuracy varies widely by language.
Looking for Wispr Flow, often typed Whisper Flow? That's a different app, a dictation tool that polishes what you say. How it compares to rawflo.
Running it on Windows
Whisper is a Python package. OpenAI's instructions are to install Python (3.8 to 3.11), then run pip install -U openai-whisper, and install the command-line tool ffmpeg, which on Windows comes through Chocolatey or Scoop. Some PCs also need Rust for one of its parts.
From there, Whisper works at the command line: you give it audio files and it writes out what was said. It transcribes recordings you already have. It doesn't type into an email or a document as you talk; that takes another program built around it.
rawflo is that kind of program, with a different speech model behind it. It doesn't run Whisper: your speech goes to Deepgram Nova-3 through rawflo's own server, and your words land where you're typing about a third of a second after you stop.
Side by side
| Whisper, run yourself | rawflo | |
|---|---|---|
| Speech to text | OpenAI's Whisper, on your own PC | Deepgram Nova-3 |
| What it does | Transcribes audio files from the command line | Types what you say at your cursor, in any app |
| Setting it up | Python, ffmpeg, then pip install; the larger models want a graphics card with several GB of memory | Install the app, sign in on rawflo.com |
| Your words | Whisper's transcript of the file | Exactly what the speech model hears, nothing rewritten |
| One recording | A file of any length you give it | Up to an hour, waiting through 2 minutes of quiet |
| The practical bits | ||
| Cost | Free and open source (MIT License) | $5 US a month with 5 hours, then $1 US an hour up to a limit you set |
| Runs on | Your own PC | Windows 10 and 11, with an internet connection |
| Languages | Many, with accuracy varying by language, OpenAI says | English (Australia, US, UK, New Zealand, India) |
Which to use
Run Whisper yourself
If you're transcribing recordings you already have, you're at home at a command line, you want it free, or your audio isn't in English. It's a strong model, and it's yours to run.
Use rawflo to just dictate
Nothing to install but the app. Press a key, talk, pause to think, press it again: your exact words land where you're typing.
Start free