Whisper turns speech into text as well as any paid service — and it runs entirely offline on your machine, free, private, even on a laptop. Here's the fastest way to set it up.
The fastest way is faster-whisper with the large-v3-turbo model: pip install faster-whisper, add ffmpeg, and you're transcribing audio to text on your own machine — offline, free, and private. OpenAI released Whisper as open source, so the models download to your computer and no audio is ever sent to OpenAI or anyone else. It's as accurate as any paid transcription service, needs only about 6GB of VRAM (or 1.5-4GB quantized), and even runs on CPU if you don't have a GPU. Here's the complete setup, with the model to pick.
Step 1: pick your setup
There are three good ways to run Whisper, depending on your hardware. faster-whisper is the pick for anyone with an NVIDIA GPU — it's a reimplementation that delivers the same accuracy as original Whisper but runs about 4x faster on GPU and 2x faster on CPU, and it's a simple pip install. whisper.cpp is best for CPU-only systems or Macs — it's written in C++ with no heavy dependencies and installs via Homebrew on Mac; it's the go-to when you don't have a capable GPU. And the original OpenAI Whisper (pip install openai-whisper) works too, with mlx-whisper being the fast option on Apple Silicon. For most people with an NVIDIA card, faster-whisper is the answer.
# fastest setup (NVIDIA GPU): faster-whisper
pip install faster-whisper
# (install ffmpeg too — e.g. `brew install ffmpeg` or your package manager)
# then in Python:
from faster_whisper import WhisperModel
model = WhisperModel('large-v3-turbo') # ~6GB VRAM, near-top accuracy
segments, info = model.transcribe('audio.mp3')
for s in segments: print(s.text)
Step 2: choose the right model
Whisper comes in sizes from tiny to large, but in 2026 the choice is easy: large-v3-turbo is the answer for almost everyone. It keeps large-v3's encoder but cuts the decoder from 32 layers to 4, delivering near-identical accuracy at a fraction of the compute — so you get top-tier transcription without needing a big GPU. It uses about 6GB of VRAM in FP16, and with INT8 or 4-bit quantization (which faster-whisper and whisper.cpp both support) you can squeeze the large model into roughly 1.5-4GB — small enough for almost any GPU, or CPU. Only drop to the smaller small or base models if you specifically need maximum speed and can accept lower accuracy. For quality transcription, large-v3-turbo is the sweet spot, and it's the same quantization principle that makes text LLMs fit modest hardware.
large-v3-turbo gives near-top accuracy at ~6GB VRAM — and quantized, it fits 1.5-4GB or runs on CPU. · Unsplash
Quick answers
How do I run Whisper locally for free?
Install faster-whisper (pip install faster-whisper) plus ffmpeg, then load the large-v3-turbo model and transcribe your audio file — it runs entirely on your machine, free and offline. OpenAI released Whisper as open source, so the models download locally and no audio is sent anywhere. If you don't have an NVIDIA GPU, use whisper.cpp instead, which runs well on CPU-only systems and Macs. Whisper's accuracy matches paid transcription services, at no cost and with full privacy.
Which Whisper model should I use?
large-v3-turbo is the best choice for almost everyone in 2026. It keeps large-v3's accuracy but cuts the decoder from 32 layers to 4, so it's much faster while delivering near-identical transcription quality. It needs about 6GB of VRAM in FP16, or 1.5-4GB with INT8/4-bit quantization. Only use the smaller 'small' or 'base' models if you need maximum speed and can tolerate lower accuracy. For quality transcription on modest hardware, large-v3-turbo is the sweet spot.
Can I run Whisper without a GPU?
Yes. whisper.cpp is designed for CPU-only systems — it's written in C++ with no heavy dependencies and runs Whisper efficiently on a normal CPU, including Macs (installable via Homebrew). faster-whisper also runs on CPU, about 2x faster than the original. Transcription will be slower than on a GPU, but it works, and with a quantized model (INT8/4-bit) even the large model fits in modest memory. So you can transcribe audio locally and free even on a laptop without a dedicated graphics card.
Whisper gives you free, private, offline transcription that rivals any paid service — faster-whisper with large-v3-turbo is the setup for most. It pairs well with the rest of a local AI setup; size any model in the VRAM calculator. Source: Codersera and OpenAI Whisper.
