Installing Whisper: Step-by-Step Guide
Whisper is one of the most powerful tools for automatic speech recognition. Developed by OpenAI, it’s free to use, supports multiple languages, and works impressively well even with noisy recordings. The current flagship is large-v3, and in 2024 OpenAI added a turbo model that is roughly 8x faster with only minor accuracy loss. But how do you actually install it?
In this guide, you’ll get clear, easy-to-follow instructions on how to install Whisper on Windows, macOS, or Linux — no programming experience required.
Preparation: What You Need Before Installing Whisper
Before installing Whisper, a bit of setup is needed. Don’t worry — no coding skills required, just a bit of patience.
Install Python
Whisper runs on Python. Recommended versions: 3.8 to 3.11.
- Download it from python.org
- On Windows, make sure to check “Add Python to PATH” during installation
- Verify installation:
python --version
Set Up a Virtual Environment (Recommended)
A virtual environment keeps the Whisper setup isolated and clean.
# Windows or macOS/Linux
python -m venv whisper-env
# Activate the environment
# macOS/Linux
source whisper-env/bin/activate
# Windows
whisper-env\Scripts\activate.bat
Install FFmpeg
Whisper relies on FFmpeg to handle audio formats.
- Windows (using Chocolatey):
choco install ffmpeg
- macOS (using Homebrew):
brew install ffmpeg
- Linux (Debian/Ubuntu):
sudo apt update && sudo apt install ffmpeg
Verify installation:
ffmpeg -version
Installing Whisper: Step by Step
Step 1: Install Whisper via pip
With Python and FFmpeg set up, install Whisper:
pip install -U openai-whisper
If you run into errors (e.g., related to tiktoken or Rust), try:
pip install --upgrade pip
You may also need the Rust compiler (rustup.rs) if dependencies fail to compile.
Step 2: Run a Test Transcription
Place an audio file (e.g., example.mp3) in your working directory. Then create a Python script:
import whisper
model = whisper.load_model("small")
result = model.transcribe("example.mp3")
print(result["text"])
Run it:
python transcribe.py
The model will download automatically on first use.
Prefer the command line? Whisper also ships a CLI, so you can transcribe without writing a script: whisper example.mp3 --model turbo.
Installing Whisper by Operating System
Windows
Install Python and FFmpeg
Ensure they’re in the PATH variable so you can run python and ffmpeg globally.
Activate Virtual Environment
python -m venv whisper-env
whisper-env\Scripts\activate.bat
Install Whisper and PyTorch (Optional)
For GPU acceleration:
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu118
pip install -U openai-whisper
Test Transcription
Use the script above. If errors occur, check audio format, file integrity, or try a simpler file.
macOS
Install Homebrew, Python, FFmpeg
brew install python@3.11
brew install ffmpeg
Create Virtual Environment
python3 -m venv whisper-env
source whisper-env/bin/activate
Install Whisper
pip install -U openai-whisper
Test Transcription
Same process as Windows. On M1/M2 Macs, PyTorch supports Metal acceleration (MPS), which improves speed, especially with larger models.
Linux (Debian/Ubuntu)
Install FFmpeg and Python
sudo apt update
sudo apt install ffmpeg python3 python3-pip python3-venv
Create and Activate Virtual Environment
python3 -m venv whisper-env
source whisper-env/bin/activate
Install Whisper
pip install -U openai-whisper
Optional: PyTorch with CUDA for NVIDIA GPUs
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
Run Test Script
Use the same method as above.
Whisper Model Sizes and Memory Requirements
Whisper downloads the model when first used. Choose based on your needs:
- tiny: Fastest, lower accuracy (~75 MB)
- base: Basic balance (~142 MB)
- small: Solid for general use (~244 MB)
- medium: High quality (~769 MB)
- large: Maximum accuracy, currently large-v3 (~1.55 GB)
- turbo: An optimized version of large-v3, roughly 8x faster with only minor accuracy loss (~1.6 GB). Added by OpenAI in 2024, it is a strong default for most transcription work, but it is not trained for translation.
Loading whisper.load_model("large") uses large-v3, while whisper.load_model("turbo") loads the faster turbo model.
The table below compares the parameter count, the GPU memory (VRAM) each model needs and the relative transcription speed (higher is faster), based on OpenAI's official model card:
| Model | Parameters | Required VRAM | Relative speed | Best for |
|---|---|---|---|---|
| tiny | 39 M | ~1 GB | ~10x | Quick drafts, low-power devices |
| base | 74 M | ~1 GB | ~7x | Simple, clear audio |
| small | 244 M | ~2 GB | ~4x | Everyday transcription |
| medium | 769 M | ~5 GB | ~2x | Higher accuracy needs |
| large (large-v3) | 1,550 M | ~10 GB | 1x | Maximum accuracy |
| turbo | 809 M | ~6 GB | ~8x | Near large-v3 accuracy at high speed |
Models are stored in:~/.cache/whisper
Common Errors and How to Fix Them
FFmpeg Not Found
Check if it’s in your PATH:
ffmpeg -version
“No module named 'whisper'”
Make sure your virtual environment is active before running scripts.
CUDA Not Recognized
Install PyTorch with the correct CUDA version for your system (CU118, CU121, etc.). You’ll find the correct option on PyTorch’s Get Started page.
Transcription Fails
Verify the audio file is supported and unencrypted. Formats like MP3, WAV, FLAC, M4A, OGG, and AAC usually work. Avoid DRM or variable bitrate issues.
Faster Alternatives: faster-whisper and whisper.cpp
If the standard installation feels slow or memory-hungry, two open-source projects run the same Whisper models more efficiently:
- faster-whisper: A reimplementation built on the CTranslate2 engine. It is up to 4x faster than the original openai-whisper at the same accuracy while using less memory, and installs with
pip install faster-whisper. - whisper.cpp: A lightweight C/C++ port that runs Whisper efficiently on the CPU, which is handy on machines without a dedicated GPU.
Both reuse the same underlying models, so your choice of tiny, small, medium, large-v3 or turbo still applies. For a wider overview, see our guide to the best Whisper alternatives and how to use Whisper for speech-to-text.
Conclusion: Install Whisper or Choose an Alternative?
With just a few setup steps, Whisper can run on any modern computer. It’s a robust tool for transcription, podcasts, research, or content creation, fully offline and free.
If you'd prefer a no-setup experience for meetings, a done-for-you assistant like Sally records and transcribes your calls on Zoom, Microsoft Teams, Google Meet and Webex, then adds AI summaries, task detection and CRM integration through a plug-and-play interface. Sally is GDPR-compliant and hosted on German servers. Whisper is the developer's toolkit; Sally is the ready-made meeting assistant alternative.
Voice recognition has never been this accessible. Want to save time? Try Sally for free today.
