37 hours of calls on Sume STT: $22.20 in 222 ten-minute chunks

Thirty-seven hours of audio is 2,220 minutes, which Sume STT 1.0 bills at $0.01 per audio minute: $22.20 in total. The catalog says a request reserves from the duration you give, with a one-minute default when you omit duration_seconds and a maximum of 10 minutes, so a plan for long recordings is 222 chunks of ten minutes, each holding up to $0.10.

This is the Sume rate as of 2026-10-09. For other speech-to-text prices, read each vendor's page; this post does not compare them.

The arithmetic by chunk size

Chunk length changes the number of requests and the worst-case hold but not the total, as long as chunks are whole minutes.

Why the duration hint matters

If you omit duration_seconds, Sume reserves one minute. A ten-minute chunk submitted without the hint reserves $0.01 instead of the $0.10 it needs, and the docs describe the reservation as the estimate. Send the hint so the reservation matches the work and a low balance fails at submit with a 402.

Minute boundaries matter. The related reference-ingest docs reserve per ceil(minute) of the hint, so a 61-second chunk reserves two minutes. Cut chunks at whole minutes, and avoid a tail of a few seconds that rounds up to a full minute.

Silent audio and other STT paths

Two other Sume paths use the same STT rate. Video inspect with transcribe: true adds $0.01 per audio minute to its reservation, with a hint maximum of 600 seconds. Reference ingest can transcribe too, with a hint of at most 300 seconds, and it settles the charge to zero when the track is silent or has no speech. For inspect, check probe.has_audio first: a silent clip returns inspect_source_has_no_audio.

For a call archive the plain STT endpoint is the right path. The other two are for video you also want to inspect.

Planning the run

Submit chunks in waves that match your plan: 24 accepted jobs on Pro means 222 chunks go in about 10 waves, and the balance held at the peak is 24 x $0.10 = $2.40. Store the job id and the chunk offset together so a failed chunk can be retried alone with the same idempotency key. A failed job refunds its hold, so a retry does not double the cost. When you stitch transcripts, keep the offset of each chunk in minutes so timestamps line up with the original recording.

What the estimate does not cover

The $22.20 is the transcription line only. Moving the audio into the workspace, any later summarising by an agent and storage of the results are separate. The catalog lists applied estimates with an agent fee for calls made through an agent: $0.02 for a one-minute STT estimate against $0.01 direct, so a pipeline that runs through an agent should budget from the applied column, not the base price.