Replace audio from 12.4 s to 14 s of a clip with Timeline parts

Seedance 2.5 lists timestamp-level editing of audio and video (ByteDance Seed, read 2026-10-05). If your goal is narrower, replacing the audio between 12.4 and 14 seconds of a finished clip, you do not need a video model for it. On Sume you detach the clip's audio, then build a new audio spine from three parts (before, replacement, after) and render the clip over it with Timeline 1.0. The ffmpeg-only media routes involve no provider inference.

The four steps

All URLs must already be on media.sume.com for your workspace. Import files first with POST /v1/media-imports, and send an Idempotency-Key on every write.

• Detach the clip's audio to a wav with POST /v1/audio-detach (default sample-exact wav). Public rate: $0.01 per job.
• Produce the replacement line (for example with a text-to-speech model) as a Sume-hosted wav about 1.6 seconds long.
• Render with Timeline 1.0 using audio.parts[]: slice 0 to 12.4 from the original, then the replacement, then the original from 14 on.
• Listen at the two seams. Wav parts join in the sample domain with no gap, so any click comes from the content, not the join.

The render call

audio.parts[] takes up to 20 gapless slices, each with a url and optional source_in and duration. The video slot uses the original clip for the full 30 seconds. The three part lengths must add up to at least audio.duration_seconds, or the render is refused with audio_parts_shorter_than_duration.

curl -X POST https://api.sume.com/v1/timeline-1.0/render \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: audio-swap-001" \
  -d '{
    "audio": {
      "duration_seconds": 30,
      "parts": [
        { "url": "https://media.sume.com/artifacts/artf_demo/orig.wav", "source_in": 0, "duration": 12.4 },
        { "url": "https://media.sume.com/artifacts/artf_demo/new-line.wav", "duration": 1.6 },
        { "url": "https://media.sume.com/artifacts/artf_demo/orig.wav", "source_in": 14, "duration": 16 }
      ]
    },
    "video": [
      { "source_url": "https://media.sume.com/artifacts/artf_demo/clip.mp4", "start": 0, "duration": 30 }
    ]
  }'

Pitfalls with parts and seams

Three things go wrong most often. First, a replacement line that is longer or shorter than the gap shifts everything after it, so measure the replacement and adjust the third part's source_in to match. Second, a stereo original joined with a mono replacement is refused for channel layout; the audio join docs list audio_parts_channel_mismatch. Detach both as the same channel layout, for example channels: "mono". Third, mp3 adds priming padding at each edge, so keep wav for anything you will join.

• Match channel layouts across parts.
• Keep wav for joins.
• Measure the replacement; the spine length is authoritative.

Why not regenerate the whole clip

A new generation produces a new take, with new motion and new faces. If the picture is approved and only a word is wrong, this audio route keeps every pixel as it was. It also costs about a dime in ffmpeg jobs rather than another model run. Use a regeneration only when the picture has to change. If the replacement line is a different voice from the original, expect a noticeable change at the seams; a short music bed under the whole clip, added as the soundtrack with duck_db, can smooth it. That field needs a real spine, which parts provide.

Where the model route still helps

If the picture must change in that window, say a new product shown at 12.4 s, audio parts are not enough. Cut the clip around the window with video trim (start plus end or duration, $0.02 per job), generate a replacement shot, and join the three clips as video[] slots. The Sume docs recommend putting the trim output into video[] with source_in 0.

Keep one rule: match the replacement's length to the gap, because Timeline coverage is declared and the spine length is authoritative.