WhisperX
Fast speech-to-text with word-level timestamps and speaker labels
Overview
WhisperX takes OpenAI's Whisper model and fixes its biggest practical gaps: no reliable word-level timing and no sense of who's speaking. It adds forced phoneme alignment for accurate timestamps, voice activity detection to cut down on hallucinated text during silence, and speaker diarization powered by pyannote-audio, all while running significantly faster than stock Whisper through batched inference. The result is a transcript you can actually use for subtitles, searchable video archives, or multi-speaker interview transcription.
It's completely free and open source under a BSD-2-Clause license, maintained by Oxford's Visual Geometry Group with over 23,000 GitHub stars behind it. The catch is that it's a Python library you install and run yourself, there's no hosted dashboard, no account, and no customer support line. Getting diarization working requires a free Hugging Face token, and running it well benefits from a GPU, so this is built for developers comfortable setting up their own environment, not people looking for a plug-and-play transcription app.
Key Features
Word-Level Timestamps: Aligns each transcribed word to its exact position in the audio, not just sentence-level timing.
Speaker Diarization: Labels who said what when multiple speakers are present, using pyannote-audio under the hood.
70x Real-Time Speed: Transcribes large-v2 Whisper models dramatically faster than the original implementation through batched inference.
Voice Activity Detection: Filters out silence and non-speech audio before transcription to cut down on hallucinated text.
Forced Phoneme Alignment: Uses phoneme-based models to tighten timestamp accuracy beyond what Whisper alone produces.
Pricing
Starting price
Free and open source (self-hosted; compute costs depend on your own hardware or cloud provider)
WhisperX: Free and open source under BSD-2-Clause license. No subscription. Requires your own compute (local GPU or cloud instance) and a free Hugging Face token for diarization features.
Disclaimer: pricing may change, confirm on the WhisperX GitHub repository before relying on any usage details.
Pros
Completely Free: Open source with no subscription, paywall, or usage cap.
Accurate Word-Level Timing: Far more precise timestamps than stock Whisper, useful for subtitles and searchable transcripts.
Built-In Speaker Diarization: Labels different speakers automatically, which most free transcription tools don't offer.
Genuinely Fast: Up to 70x real-time on large-v2, meaningfully quicker than the original Whisper.
Cons
Requires Setup: No hosted app, you need Python, a GPU for good performance, and comfort with the command line.
No Support Line: As an open-source research project, help comes from GitHub issues, not a paid support team.
Hugging Face Token Needed: Speaker diarization requires signing up for a free Hugging Face account and accepting model terms.
Not Beginner-Friendly: Non-developers looking for a simple transcription tool will find this far more technical than they need.
What Makes It Unique
Word-Level Accuracy Plus Diarization, Free: Combines precise word-level timestamps with speaker labeling in one open-source pipeline, a combination most paid transcription APIs charge extra for.
Kay Score
8
/ 10
Tool Information
Pricing
Free and open source (self-hosted; compute costs depend on your own hardware or cloud provider)
Category
Voice & Video
Platform
Web / iOS / Android
Last Updated
