Transcribe a podcast back catalogue: 100 episodes cost $31

Transcribing a 30-minute podcast episode on Sume costs $0.31 at list rates, so a 100-episode back catalogue is $31.00. That is three 10-minute speech-to-text jobs at $0.01 an audio minute ($0.30) plus one Timeline audio split job at $0.01 that cuts the episode into three windows. Rates come from Sume's pricing page and the Timeline audio docs, read 2026-10-03; STT takes at most 10 minutes per request, which is why the split is needed.

What does the catalogue cost?

Add one cent per episode for a detach if your archive is video rather than audio. The four-episode row is a month of a weekly show; the 52-episode row is a year.

How should I run a batch this size?

Give every job a stable idempotency key built from the episode and window, for example ep-0042-w2, so a restarted script cannot bill the same window twice. Reusing a key with a different payload returns 409 idempotency_conflict, which is a useful alarm that your numbering broke.

Do not fire 400 jobs at once. The generation admission docs say valid jobs queue while concurrency is full, but a full queue returns 429 queue_full. Submit in waves sized to your plan's accepted job capacity and poll GET /v1/jobs/:id/status.

What does this leave out?

• Whether STT jobs count against generation concurrency the same way as media generation is not stated in the pages I read, so start with a small wave and watch for queue_full.
• Fixed 10-minute windows cut across sentences; merge transcripts by adding each window's start to its word times.
• Accuracy on your audio is untested. Spot-check names and numbers on a few episodes before you run the full catalogue.