Long video to Instagram Reels: pick cuts from sentence segments

To turn a long video into Instagram Reels with Sume, transcribe it with sentence segmentation, pick ranges of about 15 to 60 seconds that start and end on sentence boundaries, and cut each with video trim. The transcript hint is limited to 10 minutes per call, so a long recording is cut into parts of at most 10 minutes first.

Instagram's Reels guide, read on 2026-10-02, calls 15 to 60 seconds the sweet spot for tutorials and day-in-my-life clips and 3 to 15 seconds right for punchy one-liners. The tool facts come from Sume's video inspect and video trim. For a worked highlight reel with captions, see the webinar post.

Which limits shape the plan?

Four limits decide how you cut the source. The table lists them with where they come from.

How do I get sentence boundaries?

Call POST /v1/video-inspect with transcribe: true and segmentation: { mode: "sentence" }. The result carries transcript with text, words[] and gapless sentence segments[], which the docs describe as caption-line shaped. An optional silence_split_seconds (0.2 to 3) splits on pauses. Probe and stills are unbilled; transcription is $0.01 per audio minute.

Pass duration_seconds as the hint when you know it, since omitting it reserves one minute. Pass language_code such as en if auto-detect picks wrong. If the file has no audio, the call returns inspect_source_has_no_audio, so a probe-only call first is cheap insurance.

curl -X POST https://api.sume.com/v1/video-inspect \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: talk-part-1-transcript" \
  -d '{"video_url":"https://media.sume.com/artifacts/artf_demo/talk-part-1.mp4",
       "frames":false,"transcribe":true,"duration_seconds":600,
       "segmentation":{"mode":"sentence"}}'

How do I choose the ranges?

Sume does not choose them for you: the transcript says what was said, not what is worth posting. The practical rule is to find a sentence that states a claim, answer or step on its own, start the cut at that sentence's start, and end where a sentence ends and the thought resolves.

Check each range against the lengths above. A range under 3 seconds fails the Reels API minimum; one over 60 seconds is outside the guide's tutorial sweet spot but still valid. The first second matters more than the rest: start on the claim, not on 'so, as I was saying'.

How do I cut and finish them?

Send one trim per range with a unique Idempotency-Key, using start and duration. Keep the default exact precision so the cut lands on the sentence boundary rather than up to a GOP early. At $0.02 per job, ten Reels from one talk is $0.20 in trim plus the transcript.

A 16:9 recording is not a Reel as it stands. The fit options in Timeline 1.0 (cover, contain, stretch, blur) place a landscape clip in a 1080x1920 frame, and video captions can burn in the words for sound-off viewing. Edits' own guide notes captions help with sound-off viewing too.

What is the catch?

Cropping a talking head to 9:16 with cover can cut off a person standing to one side, and Sume does not track faces in this flow. Look at stills of the result, with a free inspect call, before you publish. Sume prepares the files; picking the moments and posting them stays with you.

What does the full run cost?

Take a 30-minute talk. Cut it into three 10-minute parts at $0.02 each ($0.06), transcribe each part at $0.01 per audio minute ($0.30), then trim twelve Reels at $0.02 ($0.24). That is about $0.60 before captions or assembly, and the arithmetic uses only the rates in the docs, which say to confirm live prices in GET /v1/catalog.