What WhisperX actually adds
| Whisper | WhisperX | |
|---|---|---|
| Transcript | Yes | Yes |
| Segment timestamps | Approximate | Approximate |
| Word-level timestamps | No | Yes, via forced alignment |
| Speaker labels | No | Yes, via a separate model |
| Interface | Command line | Command line |
| Cost | Free | Free |
If you are subtitling, the word-level timing is the point. If you are transcribing interviews, the speaker labels are.
Installing on Windows
There is no installer. The path is:
# 1. A CUDA-matched PyTorch first, or you will silently land on CPU pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu121 # 2. WhisperX itself pip install whisperx # 3. Transcribe with speaker labels whisperx audio.mp3 --model large-v3 --diarize --hf_token YOUR_TOKEN
Step 1 is the one people skip. pip install torch on its own often fetches the CPU build, everything then appears to work, and transcription is five to fifteen times slower than it should be. Check with torch.cuda.is_available() before blaming the model.
The three things that go wrong
The CPU-build trap
A CPU-only PyTorch installs cleanly and runs. Nothing errors — it is just slow. Always verify CUDA is visible.
The gated model
Diarization needs a Hugging Face account, acceptance of the model terms, and a token. Without it, --diarize fails at the last step, after transcription has already run.
Non-NVIDIA GPUs
No practical CUDA on AMD or Intel under Windows. You get the CPU path. whisper.cpp-based tools are the alternative there.
When WhisperX is the right call — and when it is not
Use WhisperX if you are comfortable in a terminal, have an NVIDIA GPU, and want word-level timing or speaker labels for free. It is genuinely the best free option for diarization, and for batch work a script beats any interface.
Use something else if the setup is the job rather than the work. For a lawyer or a researcher with a recording to transcribe this afternoon, an afternoon of dependency resolution is the wrong price. Whisper GUIs cover that case, and Scrieb runs diarization locally with no Python, no CUDA install and no Hugging Face account.
- Free, terminal, NVIDIA: WhisperX.
- Free, no terminal, text only: Buzz or WhisperDesktop.
- Paid, no terminal, speaker labels and DOCX: Scrieb.
Pricing
One licence for Mac and Windows. Speaker identification runs locally, no Python and no token. Free to try.
Frequently asked questions
- What does WhisperX add over plain Whisper?
- Two things: word-level timestamps through forced alignment, and speaker labels through a separate diarization model. Plain Whisper gives you text with rough segment timings and no idea who is speaking.
- Is there a WhisperX .exe for Windows?
- No. WhisperX is a Python package — there is no standalone executable to download. Installing it means Python, pip, and on NVIDIA machines a CUDA-matched build of PyTorch. If you were looking for a double-click installer, you want a desktop app instead.
- Does WhisperX need a Hugging Face account?
- For speaker labels, yes. The diarization model is gated: you accept the terms on Hugging Face, generate an access token, and pass it to WhisperX. Transcription alone works without one.
- Does WhisperX run on AMD or Intel GPUs on Windows?
- Not practically. It depends on PyTorch with CUDA, which is NVIDIA-only on Windows. Without an NVIDIA card it falls back to CPU, which is several times slower. whisper.cpp-based tools are the usual route for AMD and Intel.
- How long does setup take?
- Plan for an afternoon the first time if Python is not already configured — installing a CUDA-matched PyTorch, resolving dependency conflicts and getting the Hugging Face token accepted are each a step that can go wrong.
- Is there anything with WhisperX accuracy and no setup?
- Scrieb runs diarization locally with no Python, no CUDA install and no Hugging Face account, on Windows and Mac. It is paid; WhisperX is free. That is the trade.