Podcast guest clips from a video call: detach once, then split

To pull five quotable clips from a recorded video call, run audio detach once on the whole recording, then one timeline-audio split with five ranges. That is $0.01 plus $0.01, or $0.02, against 5 x $0.01 = $0.05 if you detach each range separately. Both jobs are flat-rate, ffmpeg-only, with no provider inference.

Sources: Audio detach and Timeline audio, read 2026-10-04. The audio-detach page itself recommends this order for many ranges.

The four calls

The path is import, check, detach, split. A whole-track detach is limited to 900 seconds of output, so a longer call needs a range on the detach job.

Detach once

Detach with the default wav output so the split is sample-exact. The result carries audio_url, duration_seconds and the probed source duration. If the file has no track you get a stable refusal rather than a silent file, which is why the probe step is worth the cost.

curl -X POST https://api.sume.com/v1/audio-detach \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: guest-call-detach-001" \
  -d '{
    "video_url": "https://media.sume.com/artifacts/artf_demo/guest-call.mp4",
    "format": "wav"
  }'

Split into the five quotes

The split takes up to 20 ranges, each { start, end } in seconds; end omitted means the rest of the file, and ranges may overlap. The result is a segments[] list, each with its own audio_url.

curl -X POST https://api.sume.com/v1/timeline-1.0/audio \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: guest-call-split-001" \
  -d '{
    "operation": "split",
    "url": "https://media.sume.com/artifacts/artf_demo/guest-call-audio.wav",
    "ranges": [
      { "start": 62.0, "end": 88.5 },
      { "start": 240.0, "end": 271.2 },
      { "start": 455.8, "end": 480.0 },
      { "start": 702.4, "end": 731.0 },
      { "start": 811.0, "end": 840.6 }
    ]
  }'

Where the ranges come from

Find the ranges by transcribing the call first: video-inspect with transcribe: true and segmentation.mode: "sentence" returns gapless sentence segments[] with timings. The transcript is $0.01 per audio minute, with a duration_seconds hint of at most 600. Pick sentence boundaries as range edges so no quote starts mid-word.