# Whisper
OpenAI's open-source speech recognition model. Runs locally, supports multiple languages.
## Install
```bash
uv pip install git+https://github.com/openai/whisper.git
```
## Usage
```bash
whisper recording.wav --model medium
```
## whisper.cpp (Recommended)
[whisper.cpp](https://github.com/ggerganov/whisper.cpp) is a C/C++ port that runs significantly faster than the Python version, especially on Apple Silicon (Metal) and with CUDA. For local, everyday transcription this is the way to go.
```bash
# build
git clone https://github.com/ggerganov/whisper.cpp
cd whisper.cpp && make
# download a model
bash ./models/download-ggml-model.sh base.en
# transcribe
./main -m models/ggml-base.en.bin -f recording.wav
```
<!-- TODO: mention jdf-stt project as a fast way to use whisper.cpp locally, on device, in everyday life -->
## Models
| Model | Parameters | Relative Speed | Use Case |
|-------|-----------|----------------|----------|
| tiny | 39M | ~32x | Quick drafts |
| base | 74M | ~16x | Casual use |
| small | 244M | ~6x | Good balance |
| medium | 769M | ~2x | High quality |
| large | 1550M | 1x | Best accuracy |
Larger models are slower but more accurate, especially for non-English languages.