What are word-level timestamps?

Segment timing vs word timing

A basic transcript is split into segments, each a few seconds long, with one start and end time. If you search for a word and land on a segment, you are dropped at the start of that segment and have to listen for it. With word-level timestamps, the search can take you to the word itself.

Here is the difference in practice. A segment might read [00:14:02 → 00:14:09] We moved the launch to March because QA wasn't done. With word timing, “March” carries its own time, say 00:14:05.3 to 00:14:05.7.

The extra precision matters most in long recordings. In a 90-minute interview, a sentence-level result might drop you ten or fifteen seconds before the phrase, and you still have to listen for it. Multiply that by dozens of searches a day and the difference between segment timing and word timing becomes a real amount of time spent scrubbing.

How they are produced

There are two common approaches. Some speech recognition models output word timings directly as they decode. Others transcribe first, then run a separate forced alignment step that lines up each word of the finished text with the audio waveform. Alignment tends to be tighter but adds processing time.

What they make possible

• Precise search results. A query lands on the moment a phrase was said, not the start of a paragraph.
• Text-based editing. Deleting words from the transcript can cut the matching audio and video, because the edit points are known.
• Karaoke-style captions. Highlighting each word as it is spoken needs per-word timing.
• Filler and silence removal. Finding every “um” and cutting it requires knowing exactly where it sits.
• Clean subtitle breaks. Caption lines can be split at natural word boundaries instead of mid-word.

Limits

Word timings can drift by a fraction of a second, especially around music, laughter or overlapping speech, and short words like “a” or “the” are the least reliable. If a word was misrecognized, its timing belongs to the wrong word. For frame-accurate edits, it is worth nudging cut points by eye rather than trusting the numbers blindly.

Word timings are also only as good as the transcript underneath them, so correcting a misheard word in a transcript editor may not update its timing unless the tool re-aligns it.

In MediaFind

MediaFind transcribes with word-level timestamps on your own computer, so search results point to the moment a phrase was spoken. The same timing powers Create's transcript editing: you can drop silences and filler words or keep a single speaker by editing text. When you save a transcript as SRT or VTT, the captions are timed from the transcript. See making SRT subtitles locally.

Frequently asked questions

Do SRT files contain word-level timestamps?

No. SRT stores timing per caption, not per word. Word timing is usually kept in the transcript itself or in a richer format such as JSON.

How accurate are word-level timestamps?

Usually within a fraction of a second, which is fine for search and captions. Around music, crosstalk or misheard words they can drift more.

What is forced alignment?

It is a process that takes finished text and the original audio and works out exactly when each word occurs, producing word-level timing.