faster-whisper is a free Python library that turns speech into text with OpenAI's Whisper models, rebuilt on the CTranslate2 inference engine. Its README measures it at up to 4 times faster than OpenAI's own whisper package for the same accuracy while using less memory, and 8-bit mode cuts memory further on both CPU and GPU. A few lines of Python give you a transcript of an audio or video file, split into timed segments.
It earns its place as the engine inside other transcription tools, including WhisperX and the Buzz desktop app, so it is what you reach for when you want Whisper's accuracy at a better speed in your own script or app. It runs every Whisper model from tiny to large-v3 and turbo, gives word-level timestamps, has a built-in voice activity filter that skips silence, can batch long files for more speed, and decodes audio with PyAV, so you do not need FFmpeg installed.
faster-whisper is MIT-licensed and maintained under SYSTRAN, the machine translation company, with about 55 contributors and about 26,800 stars. The latest release is 1.2.1 from 31 October 2025, and the code still gets fixes every week or two.
- Repository: github.com/SYSTRAN/faster-whisper
- Licence: MIT (MIT License)
- Language: Python. Stars: 26.8K. Forks: 2,250. Last push: Nov 12, 2026.
- Scan: safe, Nov 16, 2026, commit 6cb639e
Who it is for
Developers building transcription into an app, researchers and journalists transcribing interviews in bulk, and creators making subtitles with a script. You need Python 3.9 or later; GPU speed needs an NVIDIA card with CUDA 12 and cuDNN 9. If you want an app instead, Buzz uses it under the hood.
Getting started
1. Install from PyPI (Python 3.9 or later)
pip install faster-whisper2. Transcribe a file on the CPU; the small model downloads from Hugging Face on the first run
python -c "from faster_whisper import WhisperModel; m = WhisperModel('small', device='cpu', compute_type='int8'); segs, info = m.transcribe('audio.mp3'); [print(s.start, s.end, s.text) for s in segs]"3. On an NVIDIA GPU under Linux, add the CUDA 12 libraries, then use device='cuda'
pip install nvidia-cublas-cu12 nvidia-cudnn-cu12==9.*Models download once from Hugging Face into its cache folder, ~/.cache/huggingface, and then run offline. With the pip NVIDIA libraries on Linux you also need the LD_LIBRARY_PATH line from the README. Our scan covers this library's source, not the ctranslate2, onnxruntime and PyAV packages pip installs alongside it.
Safety scan
We cloned SYSTRAN/faster-whisper at commit 6cb639e on Nov 16, 2026 and ran the checks described on the GitHub Tools page: credential patterns, decode-and-execute code, install-time scripts, committed binaries, risky CI workflows, every host the code talks to, known vulnerabilities in pinned dependencies, and project hygiene. A person read every hit. This is what we found.
- 0 secret hits across 40 files and about 6,300 lines, mostly Python. The scanner reports no binaries; one data file is worth naming: faster_whisper/assets/silero_vad_v6.onnx (1.2 MB), the Silero voice activity model, loaded by onnxruntime. ONNX files hold a model graph rather than pickled Python. The tests also carry five short audio clips.
- 0 pattern hits. We also looked for pickle loading, eval, subprocess calls and raw web requests in the library and found none; its only network call is huggingface_hub's snapshot_download for the model you name.
- No install hooks or scripts. pip install faster-whisper pulls ctranslate2, huggingface_hub, tokenizers, onnxruntime, av and tqdm from PyPI.
- No advisories reported, with a caveat: requirements.txt sets version ranges rather than a lockfile, so the advisory check has no exact versions to look up; pip installs the newest versions inside those ranges.
- The hosts in the code are GitHub, Hugging Face and documentation links. The built-in model names point to Systran's Hugging Face repositories, except turbo and large-v3-turbo, which come from Mobius Labs, and distil-large-v3.5 from the distil-whisper team. One workflow runs lint, tests and the PyPI release, with no pull_request_target; its actions are tagged by version rather than pinned, and releases use a stored PyPI token. No security policy, Dependabot or CodeQL.
What the scanner counted
| Check | Result |
|---|---|
| Secrets | None found. |
| Suspicious code | None found. |
| Install-time code | None: nothing runs at install beyond the package manager itself. |
| Committed binaries | None. |
| CI workflows | 1 workflow. None use pull_request_target. 1 of 1 third-party action pinned to a tag rather than a commit. |
| Network hosts | 4 distinct hosts referenced from source; most often github.com, docs.python.org, opennmt.net, huggingface.co. No URLs to bare IP addresses. |
| Known vulnerabilities | No lockfile to check: dependencies are declared as ranges, so what gets installed is whatever is current on the day. |
| Project hygiene | Has licence file, contributing guide. Missing security policy, automated dependency updates, CodeQL. |
| OpenSSF Scorecard | Not scored: the project is not in Scorecard's weekly index. |
By the numbers
| Stars | 26.8K |
|---|---|
| Forks | 2,250 |
| Contributors | 55 |
| Commits | 279 |
| Open issues | 5 |
| Open pull requests | 16 |
| Releases | 21 |
| Latest release | v1.2.1 |
| Licence | MIT |
| Main language | Python |
| Project age | 3 years |
| Last push | Nov 12, 2026 |
| Tracked files | 40 |
| Lines of code | 6,272 |
| Checkout size | 24 MB |
Lines by language: Python 4,108, JSON 1,743, Markdown 330, YAML 91.
Questions
Is faster-whisper free?
Yes. It is free and open source under the MIT licence, and the Whisper models it downloads are free too. Running it on your own computer costs nothing.
faster-whisper or OpenAI Whisper?
faster-whisper for speed and memory: the same models and accuracy, up to 4 times faster in the README's benchmark. OpenAI's whisper package is the reference version and needs PyTorch and FFmpeg.
Does faster-whisper send my audio anywhere?
No. Audio is transcribed on your machine. It contacts Hugging Face only to download a model the first time you use it.
What hardware do I need?
Any recent computer works on the CPU. In the README's test, the small model in int8 mode transcribed 13 minutes of audio in 1 minute 42 seconds on an 8-thread Intel Core i7. An NVIDIA GPU makes the large models much faster.
Do I need to know how to code?
Yes, a little Python. If you would rather click than code, Buzz is a free desktop app built on faster-whisper.
This post is part of GitHub Tools, where every repository is cloned and scanned before it is written up. The scan is a snapshot of one commit on one day; the repository has moved on since, so check it before you install.
