Text Selection To Timestamps

Use When

• An LLM has identified the best text span but should not be trusted to output precise timestamps.
• You need exact clip ranges from transcript text selections.
• The source transcript already has word-level timings.

Core Approach

1. Let the model choose text, not timestamps.
2. Normalize source words and selected words consistently.
3. Search near the expected time first if any approximate location is known.
4. Match contiguous word sequences back to the transcript.
5. Convert the recovered span into one or more keep ranges.

Inputs