A 28-minute talk: three detach ranges, three STT requests, $0.31

A 28-minute (1,680-second) video costs $0.31 to transcribe through Sume: three audio detach jobs at $0.01 each, plus 28 audio minutes of STT at $0.01 a minute. You need three ranges rather than two because audio detach allows up to 900 seconds of output but a Sume STT request tops out at 600 seconds, so the 600-second limit decides the split.

Two limits that interact

Audio detach takes one workspace video, caps the source at 1,800 seconds and the output at 900 seconds, and has an optional range. Its docs name 16 kHz mono as the STT shape. The STT catalog entry says the estimate is capped at 10 minutes. Two 840-second detaches would fit the detach limit and break the STT one, so use ranges of 600 seconds or less.

The three requests

This script only prints the three detach bodies and the totals so you can check them before sending anything. Each detach needs an Idempotency-Key and the video must already be on media.sume.com (import it first).

import json
ranges = [(0, 600), (600, 1200), (1200, 1680)]
bodies = [
    {"video_url": "https://media.sume.com/artifacts/artf_demo/talk.mp4",
     "range": {"start": a, "end": b}, "channels": "mono", "sample_rate": 16000}
    for a, b in ranges
]
for i, body in enumerate(bodies, 1):
    print("POST /v1/audio-detach", "Idempotency-Key: detach-%d" % i)
    print(json.dumps(body))
print("detach jobs:", len(bodies), "= $%.2f" % (0.01 * len(bodies)))
print("stt minutes:", sum(b - a for a, b in ranges) / 60)

If the video is longer than 30 minutes

Detach rejects a source longer than 1,800 seconds with source_duration_exceeded. A 47-minute (2,820-second) recording has to be cut first, for example with video trim at $0.02 per job. Two trims of 1,410 seconds give 2 x $0.02 = $0.04; each 1,410-second piece then needs three detach ranges of 600, 600 and 210 seconds, so 6 detaches = $0.06; STT is 47 x $0.01 = $0.47. The total is $0.04 + $0.06 + $0.47 = $0.57. All of that is job-level arithmetic from the published rates, and the live catalog is the place to confirm them.

Why wav at 16 kHz mono

The detach docs describe 16 kHz mono wav as the STT shape. Uncompressed pcm_s16le at 16,000 samples a second and one channel is 32,000 bytes a second, so a 600-second range is 19,200,000 bytes (about 19.2 MB) and the 1,680-second talk is 53.76 MB in total across three files. That is arithmetic from the stated sample format, not a measured file size; the container header adds a few dozen bytes.

Leaving sample_rate out inherits the source rate, which for a 48 kHz stereo video would be six times larger per second (48,000 x 2 channels x 2 bytes = 192,000 bytes a second against 32,000). Set channels and sample_rate explicitly when the only consumer is STT.

Order of operations

Submit the three detach jobs with different Idempotency-Key values, poll each with GET /v1/jobs/:id/status, read audio_url from GET /v1/jobs/:id/result, and send each URL to the STT request together with its own duration_seconds (600, 600 and 480). The catalog says that omitting duration_seconds reserves one minute, so setting the three values keeps each reservation equal to the audio it covers.