Remove filler words from a video by API: cut at word timestamps

The Sume docs list no filler-word operation. To remove "um" and "uh" you transcribe with POST /v1/video-inspect, find the filler entries in words[] yourself, and render the stretches between them as video[] slots of a Timeline. The cut list is your code's job.

Descript's API page names "Remove Filler Words" among the Underlord edits that "can use AI credits". Sume's version is manual and built from the Video inspect and Timeline 1.0 docs, read 2026-09-30.

How do I find the fillers?

Run video inspect with transcribe: true (and frames: false). The transcript has words[]. Your code scans them for the tokens you call filler and records each one's time range. The docs show no filler labels, so which words count as filler is yours to decide. Transcription costs $0.01 per audio minute.

How do I cut them out?

Invert the list: the clips between fillers are what you keep. Each becomes one video[] slot with source_in at the clip's in-point and duration at its length; start is where it lands on the output, and video[0].start must be 0. The audio spine must be cut the same way: extract the source audio with POST /v1/audio-detach, then list the same ranges in audio.parts[] (each part takes url, source_in and duration; at most 20 parts, joined gaplessly). Note the cap of 20 parts when you count your kept ranges.

Can I smooth the cuts?

Video slots after the first can take a transition, for example a fade of 0.25 s; more than 8 adjacent fades is refused with too_many_chained_transitions. Preflight with POST /v1/timeline-1.0/plan, which is unbilled and returns segment_count and estimated_cost_usd_micros. A Timeline may have up to 200 slots and 1800 s of output.