Find product moments in a live-commerce replay and trim them out

To pull product clips out of a live-commerce replay, transcribe the replay once with sentence segmentation, read which sentences pitch which product, then cut each range into its own MP4 with Sume's video trim. The transcript costs $0.01 per audio minute and each cut costs $0.02 per job, so a 10-minute replay with six product clips is about $0.22 before any captions.

Sume does not decide which moment is the best pitch. Video inspect returns probe facts, stills and an optional transcript; it does not return typed scenes. You or your own script pick the ranges, and Sume cuts them.

Why a live replay is a good source for ad clips?

A live sale is hours of selling with a few minutes that work as standalone ads: the demo, the price reveal, the answer to a common question. YouTube's help page on selling your own products in a live stream says streams tend to work best when they run "at least half an hour to an hour long" (read 2026-10-02), so most replays are long and mostly not ad material.

The audio is also real. TikTok's Requirements for High-Quality Videos and LIVEs (read 2026-10-02) asks for real-time speech in a LIVE and lists AI-generated voices and audio recordings as prohibited there. A clip cut from the replay keeps the host's actual voice. That page is about the LIVE itself, not about where you repost the clip, so check the rules of the destination platform before you run it as an ad.

How do you get a sentence-level transcript of the replay?

Import the replay first, because inspect and trim only read this workspace's media.sume.com files, as described in Media inputs. Then call inspect with transcribe: true and segmentation.mode: "sentence". Setting frames: false skips the stills so only the transcript half bills.

The docs cap the duration_seconds hint at 600 seconds, and the source itself must be 1800 seconds or less. For a replay longer than 10 minutes, our approach is to cut it into parts of 10 minutes or less with trim first, then transcribe each part. That is a workflow choice on our side, not a documented requirement.

curl -X POST https://api.sume.com/v1/video-inspect \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: replay-part1-inspect" \
  -d '{
    "video_url": "https://media.sume.com/artifacts/artf_demo/replay-part1.mp4",
    "frames": false,
    "transcribe": true,
    "duration_seconds": 600,
    "segmentation": { "mode": "sentence" }
  }'

How do you pick the ranges?

The result carries transcript.segments[] as gapless sentence segments. Search them for the product name or SKU the host says, then widen the range by a second or two on each side so the cut does not clip a word. Because the segments are gapless, the end of one sentence is the start of the next, which makes the padding easy to reason about.

A product usually comes up more than once in a long live. Take the pitch you want, not the first mention. Our rule of thumb is one product, one claim, one clip, under 30 seconds, but that is an editing preference rather than a platform rule.

How do you cut each range?

Send one trim job per range. start is in seconds and you pass exactly one of end or duration. The default precision of exact re-encodes so the cut is frame-accurate; keyframe copies the stream and can start up to a GOP early, so read actual_start_seconds in the result if you use it. The default audio is keep, which is what you want here.

Use a different Idempotency-Key for each range, and the same key only when you retry the same range.

curl -X POST https://api.sume.com/v1/video-trim \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: replay-part1-clip-serum" \
  -d '{
    "video_url": "https://media.sume.com/artifacts/artf_demo/replay-part1.mp4",
    "start": 212.4,
    "duration": 24,
    "output": { "width": 1080, "height": 1920, "fps": 30 }
  }'

What does the whole pass cost and what is left to do?

Prices below are the public rates in the docs; confirm them in GET /v1/catalog before a big batch.

The cut clips are plain MP4s. Burn captions on them with video captions if you want sound-off viewing, or join several into one highlight reel with Timeline 1.0. Neither step is needed to get the clips out.