whisper
Transcribe speech in dozens of languages with a model you can run entirely offline.
What it is
Whisper is a speech recognition model and command line tool released by OpenAI. You hand it an audio file, it hands back text — and it copes with accents, background noise and specialist vocabulary far better than the open source recognisers that came before it.
It is not a hosted service. You install a Python package, it downloads the model weights once, and after that everything runs locally on your own CPU or GPU. The same code ships in five sizes from tiny to large, so you can trade accuracy against speed depending on the machine you have.
It suits anyone who needs transcripts of meetings or interviews, subtitles for video, or a speech-to-text step inside a pipeline they would rather not send to a paid API. Because the weights are open, people have ported the same model to C++, JavaScript and mobile, so you are not locked into Python.
Why we recommend it
The weights are genuinely open, not just the paper
OpenAI published the trained weights and the code under MIT, not only an API. You can run this forever, on your own hardware, with no account, no rate limit and no per-minute bill.
One command turns a recording into a transcript
A single invocation of the whisper command processes a whole file. Flags cover language selection, translation into English, and output as subtitles, so most jobs need no scripting at all.
Accuracy holds up on messy real-world audio
It was trained on a very large and varied set of recordings, so it handles crosstalk, strong accents and cheap microphones where older tools demanded clean studio input.
Five model sizes means it fits the hardware you have
The smaller models run comfortably on a laptop CPU while the largest needs a GPU but approaches commercial quality. You are not forced to buy a GPU to get a usable transcript.
How it scores
These are our editorial ratings, not benchmark results. We score on how the project behaves in practice — how long it takes to get running, how good the documentation is, how responsive the maintainers are, and how well it holds up over time. The licence and star count above are read from GitHub and are a snapshot.
Pros and cons
Strengths
- Runs fully offline once the model weights are downloaded
- Supports a long list of languages and can translate them into English in the same pass
- MIT licensed weights and code, with ports to C++, JavaScript and other runtimes
- Clean Python API and a simple command line tool over the same model
- Very widely used, so answers and integrations are easy to find
Trade-offs
- Needs ffmpeg installed separately, which is the most common first-run failure
- The large models want a GPU; on CPU they are too slow to be practical for long recordings
- No speaker diarization — it will not tell you who said what
- It produces a transcript, not a live service; real-time streaming is not part of the package
How to get it
One-line install
Prefer the terminal? Copy a command and you are done.
pip install openai-whisper
brew install ffmpeg
sudo apt update && sudo apt install ffmpeg
Source and documentation
At a glance
| Category | Speech AI |
|---|---|
| Licence | MIT |
| Written in | Python |
| GitHub stars | 108,825 |
| Repository | openai/whisper |
Licence and star count were read from the GitHub API when this page was generated and are a snapshot. Check the repository for the current figure and the authoritative licence text.
Other Speech AI tools
We are not affiliated with whisper or its maintainers. We list it because it is genuinely useful and genuinely free. If you plan to build a commercial product on top of it, read the licence on the repository first.