Forced alignment

Forced alignment is the process of taking an existing transcript and an audio recording and computing exactly when each word (or phoneme) was spoken, producing precise word-level timestamps.

Unlike plain recognition, alignment already knows the words — the task is only timing. That is why tools like WhisperX layer a dedicated alignment model on top of Whisper: Whisper's own timestamps are approximate, and alignment sharpens them to tens of milliseconds.

Accurate alignment is what makes karaoke-style captions, clickable transcripts that seek the audio, precise clipping and dubbing workflows possible.

Related terms