whisperX
m-bain/whisperX · Python · BSD-2-Clause
Word-level timestamps on top of Whisper-class ASR
この解説はまだ日本語がありません。英語の原文を表示しています。
WhisperX adds forced alignment and better timestamps — the missing piece when you cut video to speech or burn accurate subtitles.
スコア
導入の要点
pip install whisperx 深読み
WhisperX is forced-alignment ASR built on Whisper: batched inference reaches about 70x realtime transcription with whisper large-v2, and a wav2vec2 alignment model pins accurate word-level timestamps — the part subtitle and video pipelines actually need.
It adds the pieces raw Whisper leaves out: multispeaker ASR via pyannote speaker diarization, VAD preprocessing that reduces hallucination and enables batching with no WER degradation, and a faster-whisper backend for lower memory.
The Python API exposes each stage separately — transcribe, align, diarize — so you can compose exactly the pipeline your task needs, and the alignment approach is documented in an arXiv paper.