How to Run Whisper Locally on Your Own Computer

Yes. Whisper runs on your own computer with no API key and no usage fee. You install one implementation, download a model once, and transcribe audio from the command line, from Python, or from a desktop app. The official package is released under the MIT License, and its turbo model, the one its command example uses, needs about 6 GB of video memory ^{1 2 3}.

Your hardware decides most of the setup. An NVIDIA graphics card points to faster-whisper. A Mac points to whisper.cpp or MLX. A computer with no suitable GPU can still run whisper.cpp or faster-whisper on the processor. This guide gives the install steps, the memory and speed of each model, and the settings and fixes that matter once Whisper runs. For background on the technology itself, see our guide to Whisper AI.

Whisper runs on your own computer for free once you download a model

Once a model file is on disk, transcription runs on your own processor or graphics card, so the audio does not have to be uploaded to a service. You download the finished model weights rather than training anything yourself.

The first use of a model downloads its file. The openai-whisper package keeps model files in a cache folder, which is ~/.cache/whisper unless you choose another with the --model_dir option ⁴. Plan for disk space as well as memory. The large model in whisper.cpp takes 2.9 GiB on disk and needs about 3.9 GB of memory ⁵.

The software itself costs nothing to use. The code and the model weights are released under the MIT License ¹. The real costs are your computer's time and memory, which is why the next sections focus on hardware. If you are still choosing a model, our answer on which Whisper model to use explains the trade-off.

Pick an implementation first: openai-whisper, faster-whisper, whisper.cpp, MLX or WhisperX

Pick the implementation by your hardware and by the output you need. All of them run the same family of trained models. They differ in the engine that runs the model, and that engine decides speed, memory use and which computers they suit.

openai-whisper is the reference version, and its README gives the pip command ⁷. It is the simplest to start with, and it was the slowest in both benchmarks below. Our question on whisper.cpp versus faster-whisper covers that choice.

faster-whisper reimplements Whisper on the CTranslate2 engine. Its README says it is up to 4 times faster than openai/whisper for the same accuracy, while using less memory ⁸. That is the project's own claim, so check it on your own audio. It does not need a separate ffmpeg install, because it decodes audio with the PyAV library ⁹. Its GPU mode needs NVIDIA libraries ¹⁰, so it suits NVIDIA graphics cards. On a Mac, choose whisper.cpp or MLX instead.

whisper.cpp is a C and C++ port that runs without Python. Its README describes Apple silicon as a first-class target for running Whisper, optimized through ARM NEON, the Accelerate framework and Metal ¹¹. It also runs on CPU-only machines ¹². Its command-line example needs 16-bit WAV files, so convert other formats first ⁶.

MLX Whisper is the port for Apple's MLX framework, which is built for Apple silicon ¹³. The MLX Whisper README covers its install and command-line use.

WhisperX builds on the faster-whisper backend ¹⁴. Its README reports 70 times real-time transcription with the large-v2 model ¹⁵. Its main addition is word-level timestamps. The README says Whisper's own timestamps are per utterance, not per word, and can be inaccurate by several seconds ¹⁶. Speaker labels, which the README calls speaker diarization, are a separate step with extra setup. See speaker diarization for what that step does.

If you write Python code rather than running commands, our guide to speech recognition in Python compares the libraries.

Model size sets the memory you need and the speed you get

Model size sets two things: how much memory the run needs, and how long it takes. Start with turbo, the model the README's command examples use ². OpenAI describes it as an optimized version of large-v3 that transcribes faster with minimal degradation in accuracy ¹⁷. Use medium or large when you need translation, because turbo is not trained for it ¹⁸.

The table lists each model's video memory and its speed relative to the large model. The README warns that real speed depends on language, speaking rate and hardware, so treat the speed column as a ratio, not a promise.

The .en models are English only. The README says they perform better for tiny and base, and that the difference becomes less significant for small and medium ( OpenAI's model table).

Speed and memory are only half the decision. Accuracy also depends on the language ²⁹. To compare models on your own recordings, measure the word error rate, and see our guide to how accurate speech recognition is for typical figures.

A GPU makes the large models practical, but whisper.cpp runs well on CPUs and Apple Silicon

A GPU is the fastest route for the large models, but you do not need one. In the faster-whisper README's CPU benchmark, whisper.cpp finished the small-model run in 125 seconds, while openai-whisper took 418 seconds on the same 13-minute recording ^{30 31 32}. faster-whisper's int8 mode was quicker still, at 102 seconds ³³.

The GPU picture is similar. In the README's GPU benchmark, openai-whisper took 143 seconds with the large-v2 model in fp16 precision ³⁴. faster-whisper in int8 took 59 seconds on the same recording ³⁵.

These benchmarks come from one project's README, on one recording and one set of hardware, so treat them as a guide to the ranking, not a promise for your machine. Our question on whether Whisper can run without a GPU covers the hardware side in more detail.

On a Mac with Apple silicon, whisper.cpp and MLX are the two native routes. The whisper.cpp README lists Metal among its optimizations for Apple silicon ¹¹. Its Core ML option moves the encoder onto the Neural Engine, which the project says can be more than three times faster than CPU-only execution ³⁶.

Quantization is the other lever for memory. Quantized whisper.cpp models require less memory and disk space, and can be processed more efficiently on some hardware ³⁷. The README does not give an accuracy cost for quantized models, so compare the output with the full model on your own audio before you rely on one.

The official Python package needs Python, ffmpeg and one pip command

Installing openai-whisper needs three things: a supported Python, ffmpeg, and one pip command. Then one line transcribes a file. The steps below are for a Mac or Windows computer with a terminal open.

1. Install Python. The README says the codebase is expected to be compatible with Python 3.8 to 3.11 and recent PyTorch versions ³⁸.
2. Install ffmpeg. openai-whisper requires the ffmpeg command-line tool on your system, in addition to the Python package ³⁹. On a Mac, run brew install ffmpeg with Homebrew. On Windows, run choco install ffmpeg with Chocolatey.
3. Install the package. Run pip install -U openai-whisper ⁷. If tiktoken has no prebuilt wheel for your platform, the README says you may need Rust ⁴⁰.
4. Transcribe a file. Run whisper audio.mp3 --model turbo. The README uses this command, which selects the turbo model ².

To get the newest code from the repository, rather than the last PyPI release, the README gives a force-reinstall command ⁴¹. Use it when you want a fix that has not reached PyPI yet.

Because openai-whisper hands decoding to ffmpeg, it reads any format that ffmpeg can decode, which is why ffmpeg is a requirement. The whisper.cpp command-line example works differently: it reads 16-bit WAV files, so convert other formats first. The README's example uses ffmpeg to convert an MP3 file into a WAV file first ⁶.

If you would rather not install anything, our guide to transcribing audio with Whisper, ChatGPT or Gemini covers hosted options, and our question on running Whisper in Google Colab covers a notebook route. Running locally keeps the audio on your computer, which matters for private recordings. Our guide to Whisper and privacy explains what that does and does not protect.

The settings that change results: language, initial prompt, temperature and voice activity detection

Four settings change the transcript more than any other: the language, an initial prompt, the temperature fallback, and voice activity detection. Set the language when you know it. Use a prompt for names and jargon. Leave the temperature and previous-text settings at their defaults unless you see repeated text. Turn on voice activity detection for recordings with long silences.

• Language. In openai-whisper, the language is unset by default and the program detects it ⁴². Set it with --language when you know the spoken language. The README's example is whisper japanese.wav --language Japanese.
• Initial prompt. The --initial_prompt option supplies text for the first window of audio. Its documented use is to give the model custom vocabulary or proper nouns, which makes those words more likely to be transcribed correctly ⁴³. Keep the prompt short and accurate, because it can also pull the output toward its own words.
• Temperature fallback. When a decode fails its quality checks, Whisper can retry with a sequence of higher temperatures ⁴⁴. The retries make a run slower, and setting a single temperature turns the fallback off.
• Previous text. --condition_on_previous_text feeds the previous output into the next window. The source says that turning it off can make the text inconsistent across windows, but makes the model less likely to get stuck in a failure loop ⁴⁵. Change it only if you see repeated or looping text.
• Voice activity detection. faster-whisper includes a voice activity detection filter based on Silero VAD. Passing vad_filter=True turns it on, and its default removes only silences longer than 2 seconds ⁴⁶. Shorter pauses stay in the audio, so the filter is conservative by design.

Test each change on a short recording with a known transcript before you run a long file. A single setting can help one recording and hurt another.

Desktop apps wrap Whisper for people who do not want a command line

If you do not want a terminal, desktop apps wrap the same engines in a window. Each has its own platforms and install route, so check the project's README before you choose.

• Buzz describes itself as transcribing and translating audio offline on a personal computer ⁴⁷. Its README lists Whisper.cpp as one of its backends ⁴⁸. Its Windows app is not code-signed, so Windows shows a warning when you install it ⁴⁹.
• Vocalinux is a free desktop app for X11 and Wayland, the display systems on Linux ⁵⁰. After a model is downloaded, it runs speech to text on your machine with whisper.cpp by default ⁵¹.

Check each project's README for whether the app transcribes files, dictates into other apps, or does both. If you want to speak into whatever app has focus, our overview of dictation software compares those tools. Yap, which we make, is an on-device dictation app for Mac and Windows, and you can see its options on the download page.

The common errors (ffmpeg missing, GPU not used, out of memory) each have a known fix

Most Whisper problems fall into a few groups. The table pairs each symptom with its usual cause and fix.

Start with one implementation, one model and one short file. Once that works, change one setting at a time and compare the results. For the other options, see our guides to open-source speech recognition and on-device speech recognition, and for the subject as a whole, the speech recognition overview.