Split a 14-minute episode into 20 clips with overlaps: 2 cents

Detach the episode's audio once for $0.01, then send a single timeline audio split with up to 20 ranges for another $0.01: 20 clips from a 14-minute-20-second (860-second) episode cost $0.02 in total. Ranges may overlap, which is how you give each clip a 1-second handle on both sides.

Everything here is from the Sume audio detach and timeline audio docs and the API catalog (read 2026-10-09). The tools only run ffmpeg on Sume's workers; neither uses provider inference.

Why 860 seconds works

Audio detach caps the source at 1,800 seconds and the output at 900 seconds. A whole track longer than 900 seconds needs a range, so 860 seconds fits as a whole-track wav. Twenty clips from it average 43 seconds each (860 / 20). With a 1-second handle on each side, the ranges overlap their neighbors by 1 second.

Range i could be { "start": 43*i - 1, "end": 43*(i+1) + 1 }, clamped at 0 for the first clip and left without an end on the last. The docs say that omitting end goes to the end of the file, and overlapping ranges are allowed.

Wav, not mp3

Keep wav (the default, sample-exact) if the clips will be trimmed or joined again. The docs note that mp3 adds priming padding at every edge, and a clip with a 1-second handle is meant to be trimmed again later.

A twenty-first clip needs a second split job, $0.01 more. Plan the 20 ranges to cover the sections you really need, and use the last range open-ended if you want the tail of the episode.

Reading the result

The split result is kind: timeline_audio with segments[], and each segment has its own audio_url. Because the job is asynchronous by default, poll GET /v1/jobs/:id/status or pass mode: "sync" to wait up to 30 seconds. Send an Idempotency-Key on both calls so a retry does not bill twice.

What to check

Before you publish the clips, compare the reported duration_seconds of each segment with the range you asked for. The handle overlap means neighboring clips share a second of audio, so do not re-join them by simple concatenation or you will repeat that second.

If the source is a video and has no audio track, the detach fails with detach_source_has_no_audio and you are not charged for the split you never submitted. A frames: false inspect tells you probe.has_audio first.

Both calls need an Idempotency-Key, and the tools are also exposed on hosted MCP as audio_detach and timeline_audio, which write with an idempotency_key. The flow there is the tool, then jobs_wait, then jobs_result.