AUDIO · Voice & Speech
Run 70x Real-Time Speech Transcription with Word-Level Timestamps & Diarization via Open-Source WhisperX
Vanilla Whisper suffers from timestamp drift, silence hallucinations and slow sequential decoding. Use Oxford’s open-source WhisperX (15k+ GitHub stars) combining batched Faster-Whisper, Wav2Vec2 phoneme forced alignment and PyAnnote speaker diarization for 70x real-time transcription with exact word-level timestamps.
CORE TOOLS
WhisperX (Oxford Open Source), Wav2Vec2 Alignment Model, PyAnnote Audio
BEST FOR
Short-video editors, podcasters, subtitle teams, researchers
REQUIREMENTS
Python 3.9+ (Nvidia GPU recommended for 70x speed, or CPU int8)
✦ WHY THIS WORKFLOW
Vanilla Whisper’s timestamps often drift by 500ms, ruining fast-paced video captions. WhisperX pins every word to its exact phoneme waveform while running 10x faster via VAD batching.
✦
Interactive Workflow Animation Preview
Step-by-Step Action Checklist
0 / 4 Completed
READY-TO-USE PROMPT TEMPLATE
I want to use open-source **WhisperX** to transcribe [audio type: e.g., a 2-host podcast / short-form videos needing word-by-word karaoke captions] on [GPU VRAM / CPU]. Provide: 1) [Optimal WhisperX CLI & Python Script]: Configure `compute_type`, `batch_size`, `vad_options`, Wav2Vec2 forced alignment, and PyAnnote speaker diarization (`min_speakers`, `max_speakers`); 2) [Word-Level Subtitle Formatter]: A Python snippet that takes WhisperX’s word-aligned JSON output and builds crisp, punchy `.srt` / `.ass` captions capped at ≤12 Chinese characters or ≤5 English words per line.