WhisperX: Fast Transcripts With Word Timestamps and Speaker Labels (GitHub, Scanned)

WhisperX takes OpenAI's Whisper and fixes the two things people complain about most: speed and timing. It runs Whisper through the faster-whisper backend in batches, which reaches around 70 times real time with the large-v2 model on a GPU, and it cuts the audio with voice activity detection first, which reduces the hallucinated text Whisper can produce over silence.

Its main trick is forced alignment. After transcribing, WhisperX aligns the text against the audio with a phoneme model (wav2vec2), so every word gets an accurate start and end time instead of Whisper's rough sentence-level timing. Add --diarize and a free Hugging Face token, and pyannote labels each line with its speaker. The output comes as SRT and VTT subtitles, JSON and plain text, which makes it a favourite for subtitles, podcast transcripts, interview notes and research.

It is BSD-2-Clause licensed, with about 24,800 stars, at v3.8.6 when scanned, and installs from PyPI as whisperx. The paper behind it was accepted at INTERSPEECH 2023.

• Repository: github.com/m-bain/whisperX
• Licence: BSD-2-Clause (BSD 2-Clause "Simplified" License)
• Language: Python. Stars: 24.8K. Forks: 2,498. Last push: Oct 18, 2026.
• Scan: safe, Oct 22, 2026, commit db8fde4

Who it is for

Video editors and creators who need accurate subtitles, podcasters and researchers transcribing interviews with several speakers, and developers who want word-level timestamps from Python.

Getting started

1. Install (add the CUDA toolkit first for GPU speed)

pip install whisperx

2. Transcribe a file with word timings

whisperx interview.wav --model large-v2 --highlight_words True

3. Add speaker labels (needs a Hugging Face read token and the pyannote model terms accepted)

whisperx interview.wav --model large-v2 --diarize --hf_token YOUR_HF_TOKEN

4. On a Mac or without a GPU

whisperx interview.wav --compute_type int8 --device cpu

Alignment models exist for English, French, German, Spanish, Italian and many other languages, picked automatically with --language. ffmpeg must be installed. Long files on CPU are slow; a GPU with 8 GB of memory handles large-v2.

Safety scan

We cloned m-bain/whisperX at commit db8fde4 on Oct 22, 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.

• A small codebase: 34 files and about 4,300 lines, nearly all Python. No secrets, no suspicious code patterns and no install hooks; the only hosts in the code are Hugging Face, PyTorch, GitHub and arXiv.
• One committed binary, whisperx/assets/pytorch_model.bin (17.7 MB), the pyannote voice activity model. PyTorch .bin files are pickle-based and can carry code, so this is worth knowing, but it is the project's own long-standing asset rather than something fetched at run time.
• 172 known advisories (7 critical) in uv.lock, the pinned environment for development. The criticals are in certifi, NLTK (unsafe pickle loading and Stanford wrapper flags WhisperX does not use) and Pillow image decoding. pip install whisperx resolves current versions within its ranges, which avoids most of them.
• 4 workflows, none using pull_request_target, with all 4 third-party actions pinned to commits. Licence present; no security policy, Dependabot or CodeQL.
• Diarization downloads pyannote models from Hugging Face with your token; audio stays on your machine.

What the scanner counted

The raw findings

Every hit the scanner wrote out, with a link to the exact line at the scanned commit. Secrets candidates are redacted.

By the numbers

Lines by language: Python 3,588, Markdown 456, YAML 150, TOML 80.

Questions

Is WhisperX free?

Yes. WhisperX is BSD-licensed and free for any use, and the Whisper, wav2vec2 and pyannote models it downloads are free too. The speaker diarization model asks you to accept its terms on Hugging Face before the first download.

How is WhisperX different from OpenAI Whisper?

It uses Whisper's models but runs them through faster-whisper in batches, so it is much faster on a GPU, and it adds what Whisper lacks: accurate word-level timestamps through phoneme alignment, voice activity detection to cut hallucinations, and speaker labels through pyannote.

Does WhisperX work on a Mac?

Yes, on the CPU. Run it with --device cpu --compute_type int8. It does not use the Apple GPU, so it is slower than on an NVIDIA card; for long files on a Mac, whisper.cpp is often faster, though it lacks WhisperX's alignment and speaker labels.

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.