Sume STT words_truncated: the 20,000-word cap and what it means

Sume's STT result caps words[] at 20,000 entries. If the cap is reached, the result also carries words_truncated: true and words_total with the full count. A 600-second job stays well under the cap, so you only see these fields on unusual audio.

What the fields say

The schema states that timings are never dropped silently. words_truncated is present only when the list was capped, and words_total is present only alongside it. The list is ordered by start, and the field is always present, even if empty.

Why it rarely fires

The job duration limit is 600 seconds. At a fast speaking rate that is a few thousand words, and a token list that includes spacing entries roughly doubles the entries. Even then it sits far below 20,000. A truncation therefore signals something odd, such as a music bed being tokenised.

Handling it in code

Check words_truncated before you trust the list for alignment. If it is true, use segments or split the audio into shorter ranges with audio detach and run each part separately. Never assume a missing flag means a bug; absence means the list is complete.