36 lectures of 25 minutes: three STT chunks each, cost vs MAI

A 25-minute lecture is 1,500 seconds, which is over the 600-second limit of a Sume STT job, so you cut it into three pieces and transcribe each. Thirty-six lectures are 54,000 seconds, or 15 hours. The transcription itself is $8.10 on MAI-Transcribe-2-Streaming and $9.00 on Sume STT, and the cutting adds about $1.08 on Sume.

The bill

The extra step is the cost of long files on Sume. Cutting is cheap, but you have to do it.

Cut with audio detach

For a lecture that exists as a video, Audio detach can cut and extract in one step. It takes a Sume-hosted video, a range of start and end seconds, and returns a new audio file. The source may be up to 1,800 seconds, so a 25-minute lecture is accepted, and each output is capped at 900 seconds, so one whole-track detach would fail. Three ranges, 0 to 600, 600 to 1,200 and 1,200 to 1,500, give three files that are each under the STT cap.

Each detach job is $0.01 at list. Three per lecture is $0.03, and 36 lectures is 108 jobs and $1.08. The docs suggest detaching once and splitting with Timeline audio when you need many ranges, but the split output also needs a Sume-hosted source and costs a flat $0.01 a job, so for three ranges the direct route is simpler.

Boundaries and the offset

Ask for 16,000 Hz mono output in the detach job. The docs describe that as the speech-to-text shape, and it keeps each chunk small. Then submit one STT job per chunk with its duration_seconds and a date-and-index Idempotency-Key.

Chunk boundaries cut words. Pick a boundary at a pause if you can find one, or overlap each range by a few seconds and discard the repeated words when you join the transcripts. The stitch step then adds each chunk's start offset to every word time.

What the table leaves out

Microsoft's figure is the launch price from 2026-10-01 (Microsoft AI, read 2026-10-05): $0.54 per hour for MAI-Transcribe-2-Streaming through year end. This post treats that hourly rate as pro rata per second of audio, which is an assumption, so check Microsoft's billing terms before you rely on the last cent.

The two are different kinds of service. MAI-Transcribe-2-Streaming is a streaming model, built to transcribe audio as it arrives. Sume STT is an async job: you send a public HTTPS audio_url, poll GET /v1/jobs/:id/status, and read text, word times and optional sentence segments from /result. A recorded file suits the job form well, and a live call does not.

Each Sume job takes at most 600 seconds of audio and each submit needs an Idempotency-Key. Sume bills each job by the second at $0.01 a minute, so the table above is the whole cost of the transcription itself.

Per-lecture arithmetic

Per lecture, the Sume cost is 1,500 seconds of transcription at $0.25 plus three detaches at $0.03, which is $0.28. That is $10.08 across 36 lectures. The MAI streaming price for the same audio is $8.10 at the launch rate, under the pro rata assumption above. The $1.98 gap is mostly the $1.08 of cutting plus the 90 cents between the two transcription rates.

Streaming transcription does not remove the work for a recorded lecture. It reads a stream, and you would still need to feed the file in real time or use whatever batch route Microsoft offers. Check the vendor page for that before you assume the two prices compare cleanly.

Keep the lecture index: lecture id, chunk number, range start, job id. A student search that returns a sentence needs the chunk's start offset to turn a chunk time into a lecture time, and storing the offset at submit time means you never have to work it out again.

Check the slides too. A lecture video has a picture track that the audio chunks ignore, so a transcript will miss anything shown but not said. Video inspect can read a video's frames if you need that, as a separate job.