Transcribe a 30-minute recording in three 600-second STT jobs: $0.33

The plan and the price

Cut the video into three 600-second ranges with POST /v1/audio-detach, then run POST /v1/stt-1.0/transcribe on each piece. Three detach jobs at $0.01 each plus 30 minutes of STT at $0.01 per minute is $0.03 + $0.30 = $0.33.

Two limits force the split. Sume STT accepts duration_seconds from 1 to 600, so one job covers at most 10 minutes. Audio detach reads a source of up to 1,800 seconds, which is exactly 30 minutes, and writes at most 900 seconds per output.

Why mp3 for the pieces

Sume's STT size guide puts a 10 MB ceiling on the audio file. A 600-second mono wav at 16 kHz and 16 bits is 32,000 bytes a second, or 19.2 MB, which is too big. Detach's mp3 option is 128 kbps, or 16,000 bytes a second, so 600 seconds is 9.6 MB and fits. Ask for format: "mp3" and channels: "mono".

Code

The video must already be a media.sume.com URL; import it first with POST /v1/media-imports. The script below submits, polls with backoff and never resubmits a paid job. Set VIDEO to your imported URL.

import asyncio, os, httpx
API = "https://api.sume.com"
VIDEO = "https://media.sume.com/artifacts/artf_demo/call.mp4"
H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}

async def run(c, path, body, key):
    r = await c.post(API + path, json=body, headers={**H, "Idempotency-Key": key})
    r.raise_for_status()
    job, delay = r.json()["request_id"], 2
    while True:
        s = (await c.get(f"{API}/v1/jobs/{job}/status", headers=H)).json()
        if s["status"] in ("failed", "canceled"): raise RuntimeError(s)
        if s.get("result_ready"): break
        await asyncio.sleep(delay); delay = min(delay * 2, 15)
    return (await c.get(f"{API}/v1/jobs/{job}/result", headers=H)).json()

async def main():
    async with httpx.AsyncClient(timeout=60) as c:
        for i in range(3):
            rng = {"start": i * 600, "end": (i + 1) * 600}
            d = await run(c, "/v1/audio-detach", {"video_url": VIDEO, "format": "mp3", "channels": "mono", "range": rng}, f"call-d{i}")
            t = await run(c, "/v1/stt-1.0/transcribe", {"audio_url": d["audio_url"], "duration_seconds": 600}, f"call-t{i}")
            print(i, t["text"][:80])

asyncio.run(main())

Stitching the transcript

Each piece returns words[] with times counted from the start of that piece. To get one timeline, add 600 seconds to every word time in piece 2 and 1,200 seconds to piece 3. A word may be cut at a boundary; if that matters, shift the ranges to a silent point or overlap them by a couple of seconds and drop the repeated words.

• Pass duration_seconds so the reservation matches the piece; omitting it reserves one minute.
• If a piece is shorter than 600 seconds, pass its real length.
• Poll the status URL; do not resubmit a paid job that looks slow.
• Keep the three job ids so you can fetch results again.

What can go wrong

Most failures are caught before you pay for the second step. Detach refuses a source that is not on the Sume media host (unsupported_media_source), a source longer than 1,800 seconds (source_duration_exceeded) and a video with no audio track (detach_source_has_no_audio). A range that ends before it starts, or runs longer than 900 seconds, returns audio_detach_range_empty.

If a recording is longer than 30 minutes, cut it first with a video trim, or detach it in 1,800-second slices, and then apply the same three-piece plan to each. The arithmetic stays linear: every extra 10 minutes is one more detach at $0.01 and one more STT job at $0.10, so $0.11.