Word-level timestamps for subtitles

Most transcription gives you a start and end time per line. Forced alignment gives every individual word its own, which is what makes a caption line safe to split, join, or retime without the rest of it drifting. This explains the difference and when it matters.

What a line-level timestamp cannot do

A speech model returns text with timings attached to chunks it decided on. Those chunks are rarely the caption lines you want: too long, broken at the wrong place, or spanning a pause you would rather cut on.

The moment you split one, you have a problem. The editor knows when the chunk started and ended and nothing about the middle, so it guesses, usually by splitting the duration in proportion to the characters. Speech is not evenly paced, so the guess is wrong, and the second half of the line sits noticeably late or early.

What forced alignment does instead

Forced alignment is a separate pass from transcription. It already knows what was said, so instead of recognising speech it locates each known word in the audio waveform. The output is a start and end time for every word.

Here that is a wav2vec2 alignment pass run after Whisper large-v3 produces the text. Splitting a line then needs no guessing: the boundary is a word boundary, and both halves keep the times the audio actually supports.

Where it shows up

• Splitting and joining caption lines without anything drifting, which is most of subtitle editing.
• Karaoke-style and word-by-word captions, which cannot be built at all from line timings.
• Tight reading-speed control, since you can rebalance lines by word count and keep the timing honest.
• Cutting on a word boundary, because you know exactly where each one begins.

Why converted subtitle files do not have it

SRT, WebVTT and SubViewer record when a line appears and disappears. None of them has anywhere to put a per-word time. Anything rebuilt from a subtitle file is therefore line-accurate at best, however it was originally produced, which is the honest limit of every format converter, this site's included.

The spaceless-script case

Japanese and Chinese are written without spaces, and alignment there works per character, so a "word" is one character. That is the correct unit for line breaking in those scripts. Korean is not the same case despite sharing a Unicode neighbourhood: it is written with spaces, and treating it as spaceless runs its words together.

Frequently asked questions