Word Timestamps Are Measurements With Boundaries

A caption editor needs two answers for every recognized word: what text to show, and when to light it up. Google Cloud Speech-to-Text V1 can return word startTime and endTime offsets when enableWordTimeOffsets is enabled. Those values are useful for a first subtitle alignment, provided the interface treats them as estimates tied to one transcription choice.

The word timestamps guide defines each offset as elapsed time from the start of the supplied audio. It describes time values in 100 millisecond increments and shows them under the words array of a recognition alternative. For example, a response might place “Bridge” between 1.100s and 1.500s. A player using the original file can seek to those offsets or highlight the word during that span. If the editor trims the front of the file, the application must account for that trim before applying offsets to the edited timeline.

The result structure matters as much as the numbers. A response can contain multiple recognition results, each with alternatives. Google says word offsets are included only for the first alternative. If an editor chooses a different alternative because it better matches the recording, it should not silently attach the first alternative’s word boundaries to the replacement words. Their spelling or segmentation may differ. Align the selected text separately, or mark its timing for review.

The examples also show why a timestamp should not be mistaken for a precise acoustic boundary: short words can receive identical start and end values. A zero-length interval is still a returned recognition result, not evidence that the word occupied no time. The documented increment gives a granularity, not a guaranteed error range. Music, overlapping speakers, edits, and transcript corrections can all make automatic highlighting look wrong; the guide does not promise frame-accurate alignment for those cases.

For a voice application, retain the original audio coordinate system and the selected transcript alternative with the offsets. Use the values to propose subtitle positions, then let an editor scrub and adjust suspect boundaries while listening. Before shipping karaoke-style highlighting, review the shortest words and every transcript change against playback. The output is a strong starting point for navigation and captions, with human review where visual timing must feel exact.