A 30-minute lecture video to text: three detach ranges, 33 cents

A lecture video of exactly 30 minutes (1,800 seconds) is the longest source Sume audio detach accepts. Cut it into three 600-second 16 kHz mono ranges at $0.01 each, transcribe each at $0.01 per audio minute, and the whole transcript costs 3 x $0.01 + 30 x $0.01 = $0.33.

A 31-minute video fails with source_duration_exceeded, because the cap is 1,800 seconds of source. Limits and prices are from the Sume audio detach docs and the API catalog (read 2026-10-09).

The three requests

Import the video into Sume media first; detach does not fetch from the open internet. Then send three detach requests that share the source, each with range, channels: "mono" and sample_rate: 16000 (the shape the docs call the STT shape). Range one is 0 to 600, range two is 600 to 1200, range three is 1200 to 1800. Output stays within the 900-second output cap, and each range is also within the 600-second STT duration hint.

Then send each audio_url to POST /v1/stt-1.0/transcribe with duration_seconds: 600, so the reservation matches the file instead of the one-minute default. Concatenate the three text fields in order, and add 0, 600 and 1200 seconds to each part's word times.

Why not one 1,800-second detach

Detach output is capped at 900 seconds, so a whole 30-minute track needs a range anyway. Two ranges of 900 seconds would cost $0.02 instead of $0.03, but STT's duration hint tops out at 600 seconds, so 600-second pieces keep each reservation exact. If you omit the hint the reservation is for one minute, which understates a 10-minute file.

Keep wav for anything you will re-use. If you only need text, the transcription is the end product and the audio files are temporary.

Edge cases

A video with no audio track fails with detach_source_has_no_audio. You can check probe.has_audio first with video inspect at frames: false. A range that starts past the end of the source fails with detach_start_past_source, and an empty or over-900-second range with audio_detach_range_empty.

For the lecture's chapter markers, ask STT for segmentation.mode: "sentence" and use the segment times; that is a request option on the same job and does not change the per-minute price.

Doing it as a script

Write the three ranges as data, submit the detach jobs with distinct Idempotency-Key values such as lecture-12-detach-1, and poll each job by id. When all three are result_ready, submit the STT calls. A failure in one range then costs only that range to redo, which is $0.01 for the detach and $0.10 for the transcription.

Keep the offsets in your manifest. The time of every word in part two is the STT word time plus 600 seconds, and in part three plus 1,200 seconds.