What is Faster Whisper? How it differs from Whisper, setup, Docker, GPU requirements, Turkish support and license

You install Faster Whisper with Python, give it an audio file and get the transcribed text back. It is SYSTRAN's open-source Python library that runs OpenAI's Whisper model faster; according to the README it is up to 4x faster than Whisper at the same accuracy, uses less memory and also runs on a CPU. The software is free (MIT); the computer and maintenance are on you. It has no windowed interface; you use it with Python or look at interfaces others have built. There is no separate measurement for Turkish; the model is Whisper's.

What it does: Transcribes audio faster than Whisper
Is it free: Yes, MIT license
Speed: Up to 4x faster than Whisper (README)
Interface: None; Python library
Requirements: Python 3.9+, no FFmpeg needed
Latest version: v1.2.1 (October 31, 2025)
Price: Software $0 · you provide the computer

What does Faster Whisper do? Can I use it without coding?

Faster Whisper is a Python library (code package) that runs Whisper, the model that turns audio recordings into text, with less waiting and less memory. As its transcription engine it uses a fast inference engine called CTranslate2. The model is identical to Whisper's; the speed difference comes from how it is run.

Transcribe audio with PythonWhat you get: Segment list (start, end, text)Needs a separate setup Run faster and lighter than WhisperWhat you get: Same output, in less timeNeeds a separate setup Word timing and silence filterWhat you get: Start and end second for each wordNeeds a separate setup

Who it helps: developers who need to transcribe meeting, podcast or video recordings and technical teams setting up a server. It makes sense if you will process many recordings, that is, if speed matters to you. If you only transcribe a recording now and then, Whisper or a ready-made service is less hassle.

Can I use it without coding?

Faster Whisper itself has no interface. The README lists other people's projects built on it: aTrain (windowed interface, Windows Store and Linux), Faster-Whisper-Transcriber (simple interface), whisper-standalone-win (standalone command-line program) and OpenAI-compatible servers (speaches, Whisper-FastAPI). These are not SYSTRAN's; you need to vet their setup and security separately.

What it isn't for

It does not separate speakers. For that, the README points to other projects such as WhisperX and whisper-diarize. It does not generate audio; it only transcribes.

Source: faster-whisper README (including the Community integrations section), October 15, 2026. For community projects, only the descriptions in the README are reproduced.

How to install Faster Whisper: GPU, Docker and hardware

Installation is a single command, but if you'll use a graphics card you need to install the NVIDIA libraries separately. You don't need to install FFmpeg: audio files are decoded with the PyAV package that comes with the library (a difference from Whisper).

With Python 3.9 or later› pip install faster-whisper

The latest version on PyPI is 1.2.1 (October 31, 2025). Model files download automatically from the Hugging Face Hub on first run; internet access is needed the first time. You choose the folder with download_root and offline use with local_files_only=True.

Install the latest code from the repository (including changes after the release)

› pip install --force-reinstall "faster-whisper @ https://github.com/SYSTRAN/faster-whisper/archive/refs/heads/master.tar.gz"

Running on a GPU requires two NVIDIA libraries: cuBLAS (CUDA 12) and cuDNN 9 (CUDA 12). Recent ctranslate2 versions support only CUDA 12 and cuDNN 9. If you want to stay on older CUDA 11 or cuDNN 8, the README recommends downgrading ctranslate2.

Downgrade to ctranslate2 3.24.0 for CUDA 11 + cuDNN 8, or to ctranslate2 4.4.0 for CUDA 12 + cuDNN 8.
The recommended installation route is the one in NVIDIA's own documentation.
On Linux with pip: install the nvidia-cublas-cu12 and nvidia-cudnn-cu12 packages and set LD_LIBRARY_PATH.
The README mentions a library archive from the third party Purfview for Windows and Linux.

Linux, NVIDIA libraries via pip (README)› pip install nvidia-cublas-cu12 nvidia-cudnn-cu12==9.*

The README gives no list of operating systems. It says nothing about macOS, AMD graphics cards or Apple silicon, so whether it works there is not stated in the official documentation.

The README gives no step-by-step Docker instructions; it points to this NVIDIA image so the GPU libraries come preinstalled:

The NVIDIA CUDA image from the README› nvidia/cuda:12.3.2-cudnn9-runtime-ubuntu22.04

The repository also has a docker folder (a Dockerfile and a sample script). If you want a ready-made web server (an OpenAI-compatible API), look at the community projects the README lists, such as speaches and Whisper-FastAPI; these are not SYSTRAN's.

What hardware?

The README does not list the memory needed per model size. The only measurement: the large-v2 model, an 8 GB RTX 3070 Ti, 13 minutes of audio; 4525 MB of GPU memory with fp16, 2926 MB with int8.
It also runs on a processor (CPU): device="cpu", compute_type="int8". Measurement: i7-12700K, small model, int8, 13 minutes of audio in 1 minute 42 seconds with 1477 MB of memory.
How long it takes on your own computer or a rented server depends on the hardware; server cost is not stated in the official documentation.

Sources: faster-whisper README (Requirements, GPU, Installation, Benchmark), PyPI and GitHub releases, October 15, 2026. The measurements in the README are for version v1.1.0.

How to use Faster Whisper: Python example, models and subtitles

Faster Whisper is used with Python code, not a command. Create a file and run the example below. If you don't pass a language code, it detects the language itself.

Install Python 3.9+ and run pip install faster-whisper.
Paste the example below into a .py file; replace recording.mp3 with your own file name.
Run it. The first time, the model is downloaded, so it takes a while.
Check the output by listening; the text may contain errors.

On CPU, Turkish audio file (adapted from the README example)

› from faster_whisper import WhisperModel

model = WhisperModel("small", device="cpu", compute_type="int8")
segments, info = model.transcribe("recording.mp3", beam_size=5, language="tr")

for segment in segments:
    print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))

On an NVIDIA graphics card (README example)› model = WhisperModel("large-v3", device="cuda", compute_type="float16")

segments is a generator: transcription only runs when you loop over it or call list(segments). If you don't loop, the program ends immediately and it looks as if nothing happened.

Word-level timestamps

› segments, _ = model.transcribe("recording.mp3", word_timestamps=True)
for segment in segments:
    for word in segment.words:
        print("[%.2fs -> %.2fs] %s" % (word.start, word.end, word.word))

Filter out silent parts (Silero VAD; by default only silences longer than 2 s)› segments, _ = model.transcribe("recording.mp3", vad_filter=True)

Batched transcription (README example)

› from faster_whisper import WhisperModel, BatchedInferencePipeline

model = WhisperModel("turbo", device="cuda", compute_type="float16")
batched_model = BatchedInferencePipeline(model=model)
segments, info = batched_model.transcribe("audio.mp3", batch_size=16)

Which model names are available?

Built-in names in the source code: tiny, base, small, medium, large-v1, large-v2, large-v3 (the name large points to it), turbo (large-v3-turbo), distil-large-v2, distil-large-v3 and distil-large-v3.5; tiny, base, small and medium also have English-only .en versions. When you write a model name, the automatically downloaded CTranslate2 version is used.

Your own model

With the ct2-transformers-converter tool you convert a Hugging Face Transformers-compatible Whisper model (including your own fine-tuned one) and load it by folder name.

Source: faster-whisper README (Usage, Batched Transcription, Word-level timestamps, VAD filter, Model conversion) and faster_whisper/utils.py, October 15, 2026. The Turkish example is an adaptation of the language parameter in the README; the README does not give a Turkish example.

Is Faster Whisper free? Cost and updates

Short answer: the software is free, the machine you run it on is not. The code is MIT-licensed; downloading it, using it and using it for commercial work are free. But you cover the computer, the graphics card, the electricity and the maintenance of the setup; updating when a new version comes out is also your job. There is no membership, plan or API fee.

Current version

The latest release on GitHub is v1.2.1 (October 31, 2025; upgrade to Silero VAD 6 and fixes). The last code push to the repository was October 13, 2026; so the project is alive, but changes after the latest release only come with a master-branch install.

Security record

There is no published security advisory on GitHub and no SECURITY.md in the repository (October 15, 2026). This does not mean "secure"; it only means no recorded warning is visible.

Open bugs

There are 7 open issues on GitHub. Examples: text coming out as if translated in multilingual audio (#1476), very long segments in batched mode (#1396). These are user-reported issues without official confirmation.

Source: faster-whisper LICENSE, GitHub releases and issues, PyPI, October 15, 2026. Server or electricity costs are not stated in the official documentation, so no estimate is given.

Faster Whisper license, data and Turkish support

Allowed

Use the transcribed text in your work
Include the library in your own product or service, including commercial use
Modify and distribute the code

Not allowed

Remove the license text and the SYSTRAN copyright notice
Expect a warranty: the software is provided "as is"

Transcribing Turkish speechNot measured	The README gives no language list and does not mention Turkish. The large-v3 model card on Hugging Face carries the "tr" tag, and the language parameter takes a language code (e.g. tr). There is no Turkish accuracy measurement for Faster Whisper; the large-v3 Turkish measurement in the Whisper guide belongs to the model.
Language detectionYes	If no language is given, the model tries to detect the language from the recording; the detected language and its probability are returned in the info field (README example).
Turkish interfaceNo	No interface: it is a Python library.
Speaker separationNo	Not built in; the README lists community projects such as WhisperX and whisper-diarize.

License

The code is MIT-licensed (copyright: SYSTRAN, 2023). The Systran/faster-whisper-large-v3 model card is also MIT. The license of other models (such as distil and turbo) is stated on their own model cards; check it before use.

Where does your data go?

It runs locally: your audio file stays on your machine. As for network use, the README only describes downloading the model from the Hugging Face Hub. There is no privacy statement because it is a library, not a service; community servers and interfaces built on it should be evaluated separately. This is not legal advice.

Source: faster-whisper README and LICENSE, source code (transcribe.py), Hugging Face model API, October 15, 2026.

Faster Whisper, Whisper or whisper.cpp?

All three run the Whisper model; the difference is how they run it. The numbers below come from the benchmark tables in the faster-whisper README (13 minutes of audio, beam size 5). So this is not an independent test but the faster-whisper team's own measurement.

	openai/whisper	whisper.cpp	faster-whisper
GPU (large-v2, fp16)	2 min 23 s · 4708 MB	1 min 05 s · 4127 MB	1 min 03 s · 4525 MB; batch_size=8: 17 s · 6090 MB
CPU (small, fp32)	6 min 58 s · 2335 MB	2 min 05 s · 1049 MB	2 min 37 s · 2257 MB; batch_size=8: 1 min 06 s · 4230 MB
Needs FFmpeg	Yes (Whisper guide)	Not stated in these sources	No (PyAV comes with the package)

Open the Whisper guide (Turkish accuracy, API pricing, models)

Use the same settings when comparing speed: the README says openai/whisper's default beam size is 1 and faster-whisper's is 5, and that the thread count should be equal. In the CPU row of the table, whisper.cpp is faster than single-pass faster-whisper; "up to 4x" shows mostly in the batched GPU row.

Source: the Benchmark and "Comparing performance against other implementations" sections of the faster-whisper README, October 15, 2026. The table was measured with version v1.1.0 on an RTX 3070 Ti 8 GB and an i7-12700K.

Frequently asked questions

What is the difference between Faster Whisper and Whisper?

They run the same Whisper model. Faster Whisper reimplements it with the CTranslate2 engine and, according to the README, runs up to 4x faster at the same accuracy and uses less memory. It also does not require installing FFmpeg.

Does Faster Whisper work with Docker?

The README gives no Docker steps; for the GPU libraries it points to NVIDIA's nvidia/cuda:12.3.2-cudnn9-runtime-ubuntu22.04 image. The repository also has a docker folder. For a ready-made server, the README lists the speaches and Whisper-FastAPI projects written by others.

Which models does Faster Whisper use?

tiny, base, small, medium, large-v1/v2/v3, turbo (large-v3-turbo) and the distil-large versions come with built-in names. The model downloads from Hugging Face on first run. You can also convert and load your own model.

How do I install Faster Whisper?

On Python 3.9 or later, run pip install faster-whisper. It works on a CPU right away. If you will use an NVIDIA graphics card, you must also install the cuBLAS and cuDNN 9 libraries for CUDA 12.

Is there a Faster Whisper XXL or a version with an interface?

There is no version called XXL in the official SYSTRAN repository. The README lists other people's projects such as whisper-standalone-win, which offers a standalone executable on Windows, and the windowed aTrain; these sources do not say how XXL relates to them.

Does Faster Whisper support Turkish?

The model is Whisper's; the Hugging Face card carries the Turkish (tr) tag and you can pass the language code manually. There is no separate Turkish accuracy measurement for Faster Whisper. Check the output by listening.

Is Faster Whisper free, and can it be used commercially?

Yes, the code is MIT-licensed: free, and commercial use is allowed. You must keep the license and copyright notice. The computer and maintenance costs are on you.

Related guides

Whisper WhisperX MoneyPrinterTurbo

Azure AI Speech

CapCut

Chatterbox

Sources

Not clear in the official sources: Turkish accuracy measurement (none for faster-whisper) · support for Windows, macOS, AMD and Apple silicon (the README doesn't say) · memory needed per model size (the README gives only two measurements) · setup and security of community projects · the license of the turbo and distil models · the contents of the Docker folder. Each is covered in the relevant section along with what you can do.