faster-whisper vs whisper.cpp vs OpenAI Whisper
Three ways to run the same Whisper models: what each is, its licence, what its own README claims about speed and which to pick.
The Scriptorium team ·
All three run the same family of Whisper speech models; they differ in the software around the model. OpenAI Whisper is the original Python reference implementation. faster-whisper is a reimplementation built on CTranslate2 that its authors say is up to four times faster for the same accuracy. whisper.cpp is a C/C++ port with no dependencies that runs on phones, browsers and many kinds of hardware. All three show the MIT licence. We use faster-whisper on a CPU, so we know that one best; we have not benchmarked whisper.cpp.
Side by side
| Topic | OpenAI Whisper | faster-whisper | whisper.cpp |
|---|---|---|---|
| What it is | The original Python implementation and model release | A reimplementation using CTranslate2 | A port of the model to C and C++ |
| Licence shown | MIT (code and weights, per its README) | MIT | MIT |
| Speed statement in its own README | Lists a relative-speed column by model size | "up to 4 times faster than openai/whisper for the same accuracy while using less memory" | Calls itself high-performance inference; supports quantization to reduce memory and disk |
| Where it runs (as listed on its page) | Python; the README lists the GPU memory each model needs | Python, on CPU or GPU | macOS (Apple silicon with Metal and Core ML), iOS, Android, Linux, Windows, WebAssembly, Raspberry Pi, NVIDIA, AMD and Vulkan GPUs |
| Source we read | GitHub | GitHub | GitHub |
The speed statements are each project's own claims about its own software, not a neutral benchmark, and we do not know the hardware behind them. We found no independent comparison of the three on the same machine and cite none.
The Whisper model sizes
The OpenAI README lists six sizes. As printed there: tiny 39 M parameters (about 1 GB of video memory), base 74 M (about 1 GB), small 244 M (about 2 GB), medium 769 M (about 5 GB), large 1550 M (about 10 GB) and turbo 809 M (about 6 GB), with a relative-speed column that puts tiny at about ten times and turbo at about eight times the speed of large. The same README says the turbo model is not trained for translation and returns the original language even if you ask it to translate. These are the sizes the tools above load, each in its own file format.
Which should you use?
| If you want | Look at | Why |
|---|---|---|
| To read the reference code, or follow OpenAI's own examples | OpenAI Whisper | It is the original, and its README is the reference for the model sizes. |
| Python on a server or a CPU, with the speed gain its authors claim | faster-whisper | Its README says it is faster in Python for the same accuracy, and it is what we run. |
| An app on a phone, a Mac, a Raspberry Pi or in a browser | whisper.cpp | Its page lists those platforms. |
| To install nothing | A hosted tool | Such as the free no-account preview. The trade-off is that your file goes to someone else's server. |
What does not change
- Accuracy comes from the model and the audio. The same model on a noisy recording gives a poor transcript in any of the three. See how word error rate is measured.
- Invented text. The model cards say the models can produce text that was not spoken, and no wrapper removes that. See does Whisper make things up?
- Speaker labels are a separate step. Whisper-family models transcribe speech; telling speakers apart needs additional software.
How this site runs it
TryScriptorium runs faster-whisper on its own server, on CPUs with 8-bit integer weights, and the free no-account preview always uses the small model. We published how fast that is on our hardware in our measurements, and the licence question is answered in is Whisper free for commercial use? To run Whisper yourself, start with the local Whisper guide.
Checked on 2026-10-26. The GitHub pages of OpenAI Whisper (licence sentence, model table, turbo note), faster-whisper (MIT, the speed sentence quoted above) and whisper.cpp (MIT, the platforms and quantization claims). Not read or measured: any head-to-head benchmark, the projects' release notes, or how whisper.cpp performs on our server. Menus, prices and limits change; confirm on the vendor's page before you rely on them.