Skip to content
Sonorarium
Back to blog

How to Use Whisper for Free on Your Computer (Install, Commands)

Install OpenAI's Whisper with pip and ffmpeg, run the basic commands, pick a model for your memory, and learn when a cloud service is the easier choice.

Sonorarium Team 6 min read

Short answer: with a decent computer and no fear of the command line, Whisper is enough, it's free (MIT license) and your audio never leaves your computer. Install ffmpeg, then openai-whisper with pip; with an Nvidia card on Windows, PyTorch with CUDA first. No editor. If you'd rather not install anything, Sonorarium runs it in the cloud (the audio is uploaded): 15 free minutes when you create your account, with no card. Neither Whisper nor we separate speakers.

Yes, Sonorarium is our tool. The commands and figures in this guide come from Whisper's official README and from each program's website, all opened on November 4, 2026. Open-source projects change: when in doubt, their README wins.

What you need

  • Python. The README says the code is expected to be compatible with Python 3.8–3.11 and recent PyTorch versions. The torch page on PyPI (November 4, 2026) now asks for Python 3.11 or later, so if installation fails, try Python 3.11 (that's our advice, not the README's).
  • ffmpeg, to read audio and video.
  • GPU or CPU. Whisper uses an Nvidia GPU if PyTorch finds one; otherwise it uses the CPU. The CPU works, but it takes longer. On Windows, PyTorch only finds your Nvidia card after the PyTorch with CUDA step in the next section.

Installation

ffmpeg first. These are the README's commands, by system:

# Ubuntu or Debian
sudo apt update && sudo apt install ffmpeg

# Arch Linux
sudo pacman -S ffmpeg

# macOS with Homebrew
brew install ffmpeg

# Windows with Chocolatey
choco install ffmpeg

# Windows with Scoop
scoop install ffmpeg

On Windows, if you don't use Chocolatey or Scoop, download ffmpeg from ffmpeg.org.

If you use Windows and have an Nvidia graphics card, there is one more step before Whisper: install PyTorch with CUDA. On pytorch.org, pick your system, Pip and your CUDA version in its selector and run the command it gives you; we don't copy it here because it changes with each release. On Windows, plain pip installs a CPU-only PyTorch. To check, the page itself suggests running import torch and torch.cuda.is_available() in Python: it should return True.

Then Whisper:

pip install -U openai-whisper

Basic commands

The default model is turbo. These are the README's commands:

whisper audio.flac audio.mp3 audio.wav --model turbo
whisper japanese.wav --language Japanese

With your own file and language, it looks like this:

whisper interview.mp3 --model turbo --language English
  • --language: without it, Whisper detects the language. Setting it avoids mistakes when the start of the audio is silence or noise.
  • --output_format: txt, vtt, srt, tsv, json, jsonl or all (the default writes several files at once). It isn't in the README but in the program's own help (whisper --help). For example, --output_format srt --output_dir subtitles saves only the SRT in that folder.
  • --task translate: translates into English, nothing else. The README notes that turbo isn't trained for translation, and its example uses medium.

From Python:

import whisper

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

Models, memory and speed

This table is the README's. The relative speed was measured transcribing English on an A100, so yours will differ:

Size Parameters English-only Multilingual Required VRAM Relative speed
tiny 39 M tiny.en tiny ~1 GB ~10x
base 74 M base.en base ~1 GB ~7x
small 244 M small.en small ~2 GB ~4x
medium 769 M medium.en medium ~5 GB ~2x
large 1550 M N/A large ~10 GB 1x
turbo 809 M N/A turbo ~6 GB ~8x

turbo is an optimized version of large-v3 that, per the README, transcribes faster with minimal loss of accuracy. The .en models are for English only. If your GPU has little memory, drop to a smaller model and proofread more carefully (that's our advice, not the README's). Performance varies widely by language.

Apps with a window, if you don't want the command line

Program What it is License or price
noScribe Desktop app for Windows, macOS and Linux built on Whisper. It recognizes speakers and supports around 60 languages. Its site says everything stays on your computer Free, GPL-3.0
Buzz Transcribes offline with Whisper, exports to TXT, SRT and VTT, and installs from downloads or with pip install buzz-captions (with pip it asks for Python 3.12; on Mac it needs Apple silicon) Free, MIT license
MacWhisper Mac only. Exports to SRT, VTT, TXT and DOCX, among others Free version and Pro, €64 one-time according to its site

Two warnings from their own sites. The noScribe installer is about 3 GB, and by their account one hour of interview takes about 1 to 3 hours to transcribe, longer on older machines. And Buzz warns that its Windows installer isn't signed, so Windows will show a warning.

Typical errors

  • It fails because it can't find ffmpeg. Install it with the command above and open a new terminal so it's picked up.
  • "No module named 'setuptools_rust'". The README says that if tiktoken has no pre-built version for your platform, you may need Rust, and suggests pip install setuptools-rust.
  • "FP16 is not supported on CPU; using FP32 instead". Normal if you have no Nvidia GPU: Whisper switches precision by itself. If you do have one, it means PyTorch can't see it: check the PyTorch with CUDA step.
  • It runs out of video memory. Use a smaller model from the table.
  • It writes in a language that isn't yours. Add --language.

When it isn't worth it

  • You have no GPU. It works, but the speed table doesn't apply to your CPU and times can be long.
  • Your files are long, or there are many. Your computer stays busy while it works. For multi-hour recordings, see how to transcribe a long audio recording.
  • You need an editor with the audio next to the text. Whisper hands you text files; you correct them yourself, listening to the audio separately.
  • Several people are talking. Whisper doesn't separate them. There are some ideas in how to transcribe audio with multiple speakers.
  • You don't want to maintain an installation.

The option with nothing to install

Sonorarium uses Whisper large-v3 turbo, the turbo in the table, on cloud GPUs. You upload the file (any audio or video, up to 2 GB and 5 hours) and correct it in an editor with the audio synced to the text. You download Word (DOCX), TXT, SRT or VTT, in over 60 languages. It has 15 free minutes when you create an account, with no card; for a 60-minute meeting, the 45 that are missing are covered by the 120-minute pack for $6. Sonorarium doesn't separate speakers. With Sonorarium the audio is uploaded, and it is deleted after 7 days. The other packs are on the pricing page.

To see Whisper next to other tools, there's the comparison of programs and the free limits of each one. And if your case is Word or Google Docs, read how to transcribe an audio file in Word or Google Docs.

What if you hand the audio to an assistant instead of installing anything? What ChatGPT, Gemini and Copilot accept today is in Can ChatGPT transcribe an audio file?.

Sources

Opened on November 4, 2026: Whisper README, its command-line code, noScribe, Buzz, MacWhisper, PyTorch's install page, the torch page on PyPI and ffmpeg.