How to install OpenAI Whisper

You need Python, FFmpeg and one pip command. A GPU is optional but makes the larger models practical. Steps for Windows, macOS, Linux and Google Colab, then your first transcription and fixes for the usual errors.

By VexaScribe Editorial · Published August 15, 2026 · Rewritten November 17, 2026 · Checked against openai-whisper 20250625

Quick install

With Python and FFmpeg already installed
pip install -U openai-whisper
whisper audio.mp3 --model turbo

The package is openai-whisper. pip install whisper installs an unrelated package.

Before you start

  • Python. OpenAI tested on 3.9 and expects 3.8 to 3.11 to work. We suggest 3.11: it also suits WhisperX and the latest pyannote.audio if you add speaker labels later.
  • FFmpeg. Whisper calls it to read MP3, M4A, MP4 and other formats.
  • GPU (optional). An NVIDIA card with CUDA. The small models run fine on a CPU; large and turbo are slow without a GPU. Macs run on the CPU.
  • Disk space. Models download on first use to ~/.cache/whisper; the large model is about 3 GB.

Use a virtual environment so Whisper's PyTorch doesn't clash with other projects:

python -m venv whisper-env
# Windows: whisper-env\Scripts\activate
# macOS / Linux: source whisper-env/bin/activate

Windows

1. Python 3.11 and FFmpeg (PowerShell)
winget install Python.Python.3.11
winget install ffmpeg

Close and reopen the terminal so both are on your PATH.

2. Whisper
pip install -U openai-whisper

On Windows, pip installs a CPU-only PyTorch. For an NVIDIA GPU, replace it with a CUDA build, then check it:

3. GPU (optional)
pip install --force-reinstall torch --index-url https://download.pytorch.org/whl/cu128
python -c "import torch; print(torch.cuda.is_available())"

True means Whisper will use the GPU. If you get False, update your NVIDIA driver, or pick the build for your setup on pytorch.org.

macOS

With Homebrew
brew install python@3.11 ffmpeg
python3.11 -m venv whisper-env && source whisper-env/bin/activate
pip install -U openai-whisper

openai-whisper runs on the CPU on a Mac, Apple Silicon included; there's no CUDA. turbo and small are usable; large is slow. For speed on Apple Silicon, whisper.cpp uses the Mac's GPU through Metal.

Linux (Ubuntu, Debian)

sudo apt update && sudo apt install -y python3 python3-venv ffmpeg
python3 -m venv whisper-env && source whisper-env/bin/activate
pip install -U openai-whisper
python -c "import torch; print(torch.cuda.is_available())"

On Linux, pip's PyTorch already includes CUDA libraries, so an NVIDIA driver is all you need for the GPU. Check the driver with nvidia-smi.

Google Colab

Colab has FFmpeg installed. Choose Runtime › Change runtime type › GPU, then:

!pip install -U openai-whisper
!whisper /content/audio.mp3 --model turbo

Upload the audio with the Files panel first.

Your first transcription

Command line
whisper interview.mp3 --model turbo --language en --output_format srt

The file is written to the current folder. --output_format takes txt, srt, vtt, tsv, json or all. Leave out --language to detect it automatically.

Python
import whisper

model = whisper.load_model("turbo")
result = model.transcribe("interview.mp3")
print(result["text"])

for seg in result["segments"]:
    print(f"{seg['start']:.1f}s {seg['text']}")
To translate speech into English, use --task translate with medium or large. turbo isn't trained for translation and returns the original language.

Choosing a model

ModelParametersGPU memorySpeedEnglish-only
tiny39M~1 GB~10×tiny.en
base74M~1 GB~7×base.en
small244M~2 GB~4×small.en
medium769M~5 GB~2×medium.en
large (large-v3)1,550M~10 GB1×—
turbo809M~6 GB~8×—

Speed is relative to large, measured by OpenAI on English speech on an A100. Start with turbo; drop to small or base on a CPU; use large for translation or when turbo struggles with your language. The .en versions help most for tiny and base on English audio. More in Large-v3 vs Turbo and how accurate Whisper is.

Faster alternatives

faster-whisper

The same models on the CTranslate2 engine: up to 4 times faster with less memory, according to its authors. It doesn't need FFmpeg installed. On a GPU it needs cuBLAS for CUDA 12 and cuDNN 9.

pip install faster-whisper
Python
from faster_whisper import WhisperModel

model = WhisperModel("large-v3", device="cuda", compute_type="float16")
# CPU: WhisperModel("small", device="cpu", compute_type="int8")
segments, info = model.transcribe("interview.mp3", beam_size=5)
for s in segments:
    print(f"{s.start:.1f}s {s.text}")

whisper.cpp

A C/C++ port with no Python, good for CPUs, Apple Silicon and small devices. Build it from its repository.

WhisperX

faster-whisper plus a timestamp for every word and speaker labels. See our WhisperX guide.

Docker

A minimal CPU image:

Dockerfile
FROM python:3.11-slim
RUN apt-get update && apt-get install -y --no-install-recommends ffmpeg \
    && rm -rf /var/lib/apt/lists/*
RUN pip install --no-cache-dir -U openai-whisper
WORKDIR /audio
ENTRYPOINT ["whisper"]
Build and run
docker build -t whisper .
docker run --rm -v "$PWD":/audio whisper interview.mp3 --model small

For a GPU, start from an nvidia/cuda base image, install the NVIDIA Container Toolkit on the host and run with --gpus all.

Common errors

ffmpeg: command not found / FileNotFoundError

FFmpeg isn't installed or isn't on your PATH. Install it as above and open a new terminal.

No module named 'whisper'

Whisper was installed into a different Python or virtual environment. Activate the right one, or run python -m pip install -U openai-whisper with the same python you use to run it.

Errors building tiktoken, or "No module named setuptools_rust"

There's no prebuilt tiktoken for your platform, often because the Python version is very new. Use Python 3.11, or install Rust and run pip install setuptools-rust.

Whisper ignores the GPU

torch.cuda.is_available() returns False: on Windows, install the CUDA build of PyTorch (see above); on Linux, check the driver with nvidia-smi.

CUDA out of memory

The model is too big for the card. Use turbo (about 6 GB) or a smaller model, and close other GPU programs.

The transcript repeats or invents text

Whisper can loop or make up phrases over long silences or music. Trim silence first, or use faster-whisper with vad_filter=True.

Don't want to run it yourself?

VexaScribe's Standard tier runs OpenAI Whisper and adds speaker labels, with nothing to install. Upload audio or video up to 5 GB, then export TXT, Word, SRT, VTT or JSON.

30 free minutes on signup, no card. Plans from $2 a month.

A VexaScribe transcript with a timestamp and the speaker, Sam or Rosa, on every line

Sources

Checked November 17, 2026. Whisper is MIT-licensed, so commercial use is allowed, and it runs entirely on your machine. VexaScribe is independent and not affiliated with OpenAI. See our editorial standards.

Frequently Asked Questions

Is OpenAI Whisper free?

Yes. The code and model weights are released under the MIT licence, which allows commercial use. Running it yourself costs only your hardware and electricity.

Which Python version should I use for Whisper?

OpenAI tested Whisper on Python 3.9 and expects 3.8 to 3.11 to work. Python 3.11 is a safe choice that also suits WhisperX and the latest pyannote.audio.

Do I need a GPU to run Whisper?

No. tiny, base and small run acceptably on a CPU. For turbo and large an NVIDIA GPU makes a big difference: turbo needs about 6 GB of GPU memory and large about 10 GB.

Does Whisper send my audio to OpenAI?

No. The open-source package downloads the model once and then runs entirely on your computer. Only OpenAI's paid API sends audio to OpenAI.

Why does pip install whisper not work?

The package is called openai-whisper. pip install whisper installs an unrelated project. Run pip install -U openai-whisper instead.