Review a creator clip's transcript before reusing it as a paid ad

Before a creator or live-selling clip becomes a paid ad, get its transcript with POST /v1/video-inspect and transcribe: true ($0.01 per audio minute), then read the spoken claims against your own checklist. Sume returns the words; it does not judge whether a claim is allowed.

That boundary matters. The transcript turns a minute of audio into text you can search, so a person or your own script can spot lines to cut before the clip goes into a campaign.

Keep the transcript with the ad's brief so a reviewer can see exactly what was said in the version you shipped, not only what the caption on screen reads.

Why read the transcript at all?

Product clips made for organic posts are often loose about claims. Platform guidance for sellers warns about this: TikTok's seller content guide, which I read on 2026-10-02, tells sellers to avoid medical, weight-loss and exaggerated claims in their content. A holiday ad that reuses a creator's clip inherits whatever they said out loud.

Skimming a video by eye misses speech. A transcript is quick to scan, and because you can pair it with a trim, you can keep the good ten seconds and drop a risky sentence.

How do you request just the words?

Pass frames: false so the call is probe only for the visual half, and transcribe: true. Only the transcript is billed; the probe and stills are free. With segmentation.mode: "sentence" you also get sentence segments[], which are shaped like caption lines.

Set duration_seconds (the hint is capped at 600) to match the clip, otherwise the reservation is one minute. The call is sync by default, waits up to 30 seconds and answers 200, or 202 with a job to poll. Source clips can be up to 1800 seconds. A silent clip returns inspect_source_has_no_audio, so check probe.has_audio first if you are sweeping a folder.

Cost stays small: a 45-second clip is under one audio minute of transcript, so a batch of 20 creator clips of that length is on the order of twenty cents. Confirm the live rate in GET /v1/catalog before you budget a large folder.

What does a simple claims pass look like?

The script below sends one clip, splits transcript.text into sentences and prints any that contain a word from WATCH. That list is yours: it is a plain substring match, so it misses paraphrases and flags innocent uses. Treat a hit as a prompt for a human to read the sentence, not a verdict.

If the call returns 202, the script exits with the request id; read it later at GET /v1/video-inspect/{id}.

import json, os, re, urllib.request

KEY = os.environ["SUME_API_KEY"]
WATCH = ["cure", "guaranteed", "lose weight", "doctor", "#1", "best ever"]

body = {"video_url": "https://media.sume.com/artifacts/demo/creator.mp4",
        "frames": False, "transcribe": True, "duration_seconds": 60,
        "segmentation": {"mode": "sentence"}}
req = urllib.request.Request(
    "https://api.sume.com/v1/video-inspect", json.dumps(body).encode(),
    {"Authorization": f"Bearer {KEY}",
     "Content-Type": "application/json",
     "Idempotency-Key": "claims-pass-001"})
with urllib.request.urlopen(req) as r:
    out = json.load(r)
    status = r.status
if status != 200:
    raise SystemExit("202: poll GET /v1/video-inspect/" + out["request_id"])
text = out["video_inspect"]["transcript"]["text"]
for sentence in re.split(r"(?<=[.!?])\s+", text):
    hits = [w for w in WATCH if w in sentence.lower()]
    if hits:
        print(hits, "->", sentence)

What happens after a flagged line?

You have three options: drop the clip, ask the creator for a new take, or cut around the sentence with video trim ($0.02 per job) and join the pieces. If you keep the clip, burn captions only after the edit, so the text on screen matches what is said.

What Sume does not do here is check the claim against any rule, read the platform's policy or approve the ad. That review stays with you.