A weekly 45-minute interview transcript: 50 cents in five STT chunks

You record a 45-minute interview every week as a video. You want a timestamped transcript. Sume STT 1.0 accepts up to 600 seconds per job, and audio detach outputs up to 900 seconds per call. So the interview has to be cut into pieces of at most 600 seconds.

The arithmetic

Make the ranges

2,700 seconds divided by 600 is 4.5, so four chunks of 600 seconds and one of 300. This Python prints the five range objects for audio detach. The detach range field is {start, end} in seconds, and a range longer than 900 s is rejected.

duration = 45 * 60
chunk = 600
start = 0
ranges = []
while start < duration:
    end = min(start + chunk, duration)
    ranges.append({"start": start, "end": end})
    start = end
for r in ranges:
    print(r)

Where the cut falls

A hard cut every 10 minutes can land mid-sentence. Words near a boundary may be split between two transcripts. Either accept that for rough notes, or cut at a quiet moment you find by listening, then keep the ranges under 600 seconds.

Send each chunk

Submit each detached wav as audio_url with duration_seconds set to the chunk length so the reservation matches the work, and an Idempotency-Key per chunk such as ep-12-chunk-3. Merge the transcripts by adding the chunk's start offset to each word time. Sentence segmentation (segmentation: {mode: "sentence"}) gives you chapter-style lines.