STT duration_seconds omitted: Sume reserves one minute. What to send

If you leave duration_seconds out of a Sume transcription, Sume reserves 1 audio minute. At the public rate of $0.01 per audio minute, that hold is one cent. The maximum hint is 600 seconds. Send the real length of the audio so the reserved amount matches what you will use.

The numbers

The docs also note that each inspect reserves its Modal compute ceiling, and the transcript adds its per-minute rate to that reservation. The capture is never more than the hold.

Practical use

A silent clip does not transcribe: video inspect returns inspect_source_has_no_audio, so check probe.has_audio first.

• Take the length from source_duration_seconds on a detach result, or from probe on inspect, and pass it as duration_seconds.
• For audio longer than 600 seconds, detach in ranges of at most 900 seconds, then transcribe each piece. Each STT call has a 600-second hint maximum.
• Keep language_code as a hint only when you know the language. Without it, STT uses auto-detect.

Worked example

Take a 4-minute talking-head clip, 240 seconds. Without duration_seconds, the reservation is for 1 minute, which is less than the audio you need. Send duration_seconds: 240 and the reserved STT amount covers four minutes at $0.01 per audio minute, which is four cents. The hint is a size for the hold, so it should be your best estimate of the real audio length.

The same habit helps in a batch. When you transcribe many short pieces, for example the output of a timeline audio split, you know each range's start and end. Subtract them and pass the difference. You do not need to run probe on each piece.

Whatever you send, read the live rate in GET /v1/catalog before you budget, because the docs tell you to confirm the rate there and not to hard-code it.