What is WhisperX? Speaker diarization, Hugging Face token, setup, how it differs from Whisper, Turkish support and license

You install WhisperX with pip and pass it an audio or video recording from the command line; you get the transcript with word-level timestamps and speaker labels. It is an open-source (BSD-2-Clause) Python library and command-line tool built on top of Whisper; the README claims it runs 70x faster than real time, which is the maker's claim. The software is free; GPU or server costs and updates are on you. Speaker diarization requires a Hugging Face token and accepting a model agreement. An alignment model is defined for Turkish, but there is no quality benchmark.

What it does: Transcribes recordings with word-level timing and speaker labels
No-code: No: Python and the command line
Free: Yes, BSD-2-Clause; server costs are on you
Speaker diarization: Hugging Face token required
Hardware: Under 8 GB of GPU memory for large-v2 (maker's claim)
Turkish: Alignment model available; no quality benchmark
Current version: 3.8.6 (25.05.2026)

What does WhisperX do? Can I use it without code?

Can I use it without code?

No. WhisperX is not an app but a library installed with pip; it runs from the command line or from Python. There is no version with a graphical interface. For people without a GPU, the README links to three third-party demos on Replicate; check the Replicate page for their account requirements and pricing.

WhisperX listens to an audio or video recording and turns the speech into text, finds the second each word starts and ends, and, if you want, labels who said what in a recording with several speakers. It is useful for preparing subtitles, searching meeting or interview recordings, or editing.

Transcribe a recordingWhat you get: Speech segments and subtitle filesNeeds a separate setup Find word timingsWhat you get: Word start and end timesNeeds a separate setup Separate speakersWhat you get: Speaker labels on segments and wordsNeeds a separate setup

	OpenAI Whisper (as described in the README)	WhisperX
Timestamps	Sentence level; README: can be off by several seconds	Word level (alignment with wav2vec2)
Batched inference	Not supported natively	Yes (faster-whisper backend)
Speaker diarization	Not described in the README	Yes (pyannote-audio, with an HF token)
Silent sections	Not described in the README	VAD preprocessing; README: reduces hallucinated text

Who it helps: Developers and technical content creators comfortable with Python or the command line who produce transcripts and subtitles. If you are looking for a no-code tool, WhisperX is not for you.

How do I install WhisperX? Speaker diarization, Mac and use without a GPU

The commands come from the official README.

If you will use a GPU, install CUDA toolkit 12.8. The README links to the NVIDIA download pages for Linux and Windows. Skip this step if you will only use the CPU.
Use a Python version from 3.10 to 3.13. The package does not support Python 3.14 or later.
Install ffmpeg if needed (the README points to the OpenAI Whisper setup guide).

Install› pip install whisperx

Run without installing› uvx whisperx

Run the version on GitHub› uvx git+https://github.com/m-bain/whisperX.git

Install from source

› git clone https://github.com/m-bain/whisperX.git
cd whisperX
uv sync --all-extras --dev

README: the development version may contain experimental features and bugs; use the PyPI release for production.

Windows and Linux + NVIDIA GPUYes	The README gives Windows and Linux links for CUDA toolkit 12.8. It states under 8 GB of GPU memory for large-v2 (maker's claim).
CPU onlyYes	README: the --device cpu --compute_type int8 command. It may be slower; no speed figure is given.
Mac (Apple GPU)Unclear	For Mac, the README only gives the CPU route. Mac GPU support is an open request on GitHub (issue 109).

Hugging Face token for speaker diarization

Create a Hugging Face account and generate a token with “read” permission at huggingface.co/settings/tokens.
Accept the model agreement: Fill in the form on the huggingface.co/pyannote/speaker-diarization-community-1 page. The form asks for your company or university and intended use, and asks you to allow occasional email updates. The model is CC-BY-4.0 licensed and is described as “always freely accessible”.
Pass the token to the command: --diarize --hf_token <your-token>.

The token is tied to your account: do not put it in shared scripts or screenshots.

How do I use WhisperX? Ready-made commands and limits

The basic command is the name of the recording file. You add options at the end.

Simplest form (default whisper small)› whisperx recording.wav

Large model for more accurate timing, word-highlighted subtitles› whisperx recording.wav --model large-v2 --highlight_words True

Separate speakers (give the number if you know it)› whisperx recording.wav --model large-v2 --diarize --hf_token <your-token> --min_speakers 2 --max_speakers 2

Turkish recording (adapted from the README's German example)› whisperx recording.wav --model large-v2 --language tr

Without a GPU, on the CPU› whisperx recording.wav --compute_type int8 --device cpu

README: for languages other than English, pass the language code and use a large model. The Turkish command is adapted from the README's German (--language de) example; the WhisperX documentation has no benchmark or example for Turkish.

If GPU memory runs out: --batch_size 4, a smaller --model base or --compute_type int8 (may affect quality).
Words containing numbers and symbols, such as “2014.” or “13.60”, get no timestamps; characters missing from the alignment model's dictionary cannot be aligned (README).
Neither Whisper nor WhisperX handles overlapping speech well; speaker diarization is not perfect (README).
A language-specific wav2vec2 alignment model is required; for a language not on the list, you need to find your own model.

Is WhisperX free? Cost and current version

The software is free; server costs and updates are on you. There is no subscription or license fee. The real cost is a GPU (or a slow CPU) and electricity; the official source gives no amount.

Current version

PyPI stable release 3.8.6 (May 25, 2026), Python 3.10 to 3.13. GitHub has a 3.8.7rc1 pre-release (June 26, 2026). The repository is not archived; last updated September 26, 2026; 230 open items (issues and pull requests).

Replicate demos

For people without a GPU, the README links to three demos on Replicate (large-v3, large-v2, medium). These are third-party; check the Replicate page for pricing and account requirements.

Sponsor link

The Recall.ai meeting API at the top of the README is a sponsor promotion, not part of WhisperX.

Source: README, LICENSE, PyPI and the GitHub API, October 15, 2026. This is not legal or financial advice.

Does WhisperX work in Turkish? License and data

Turkish transcriptionUnclear	Turkish is on the language list of the underlying Whisper models; the WhisperX documentation has no figure or example for Turkish quality.
Turkish word timingPartly	The code defines a wav2vec2 alignment model for Turkish (tr) (a community model); its quality has not been measured.
Turkish speaker diarizationUnclear	The official source gives no language-specific benchmark; it states that diarization in general is not perfect.
Interface languageNo	No interface; command line and Python.

Allowed

BSD-2-Clause: commercial use, modification and distribution allowed; the copyright and warranty disclaimer notice must be kept

Not allowed

The speaker diarization model is licensed separately: CC-BY-4.0 (attribution required) and the Hugging Face user agreement
The README does not list the licenses of the Whisper, alignment and voice activity models; check each one before using them in commercial work

Where does the audio go?

The README gives no privacy statement. The library processes the recording on your computer; on first use it downloads alignment models (torchaudio or Hugging Face), and it needs a Hugging Face token for speaker diarization. The documentation says nothing about telemetry.

People in the recording

A speech recording can contain personal data. Consent from the people in the recording and compliance with data protection law (in Turkey, KVKK, the Personal Data Protection Law) are your responsibility. This is not legal advice.

Frequently asked questions

How do I do speaker diarization with WhisperX, and do I need a Hugging Face token?

Use the --diarize flag and pass a read token from Hugging Face with --hf_token. First you must accept the user agreement for the pyannote/speaker-diarization-community-1 model on Hugging Face.

What is the difference between WhisperX and Whisper?

WhisperX adds word-level timestamps, voice activity detection and speaker diarization on top of faster-whisper-based Whisper, and speeds it up with batched inference. The README claims 70x speed; that is the maker's claim.

How do I install WhisperX?

pip install whisperx or uvx whisperx. If you will use a GPU, install CUDA toolkit 12.8 first; with the CPU only, this step is not needed. Python 3.10 to 3.13.

Does WhisperX run on a Mac or without a GPU?

The README gives the --device cpu --compute_type int8 command for the CPU and notes that it also applies to Mac OS X. Mac GPU support is an open request on GitHub.

Is WhisperX free, and can it be used commercially?

Yes, it is BSD-2-Clause licensed; commercial use is allowed. The speaker diarization model is separate and CC-BY-4.0 licensed: attribution is required. This is not legal advice.

Tasks and setup kits that use WhisperX

Transcribing and summarizing a meeting

Related guides

Sources

Not clear in the official sources: Quality of Turkish transcription, alignment and speaker diarization · Mac GPU (MPS) support · RAM and disk requirements, and the conditions behind the 8 GB GPU memory and 70x speed claims · pricing of the Replicate demos · commercial terms of the Whisper, alignment and VAD models · telemetry. Each is covered in the relevant section along with what you can do.