Transcribe 500 thirty-second ad clips: MAI Streaming hours vs Sume

Five hundred ad clips of 30 seconds each are 15,000 seconds of audio, or 4.17 hours. Transcription costs $2.25 on MAI-Transcribe-2-Streaming at its launch price and $2.50 on Sume STT at 1 cent a minute. A whole ad library is under three dollars either way.

The bill

A transcript of an ad clip has uses beyond captions: a searchable claims list, a language check and a record of what each version said.

Five hundred small jobs

A 30-second clip is far inside the 600-second cap, so each is one job. Because each file is so short, the job count matters more than the audio length. Five hundred jobs means 500 submits with 500 different idempotency keys, so build the key from the ad's id and version.

Set duration_seconds: 30 on every submit. The documented range is 1 to 600, and without it Sume reserves one minute per job, so a short clip would reserve twice the audio it needs. A tight reservation matters for a batch, because 500 oversized reservations add up before any job finishes.

Run the batch politely

Poll jobs with backoff and a bounded pool, not 500 tight loops. Sume answers 429 with rate_limited or queue_full when you push too hard, and the right response is to wait and retry the same request with the same key, which is why the key matters.

A batch of 500 will finish at different times. Collect completed jobs as they land and write each transcript to your own store keyed by ad id. A failed job is retried alone, not the batch.

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.

What to do with the text

The transcripts open a few checks. Scan for claim words your compliance list bans. Compare the transcript with the script you approved, and flag clips where the spoken ad drifted. Read each result's detected language and compare it with the market you meant to ship it in.

Sentence segments help for subtitles. Request segmentation: {"mode": "sentence"} and each sentence arrives with a start and end time that you can turn into a caption file.

Store the language and the confidence that come back with each result. A clip that reads as a different language from the one you expected is either a mislabeled file or an ad that was recorded in the wrong market, and both are cheap to catch at 1 cent a minute before the campaign runs.

Do not forget the unit economics of the pilot. Ten clips cost five cents on Sume, so run ten first, read every transcript against the script, and fix your submit code before you launch the other 490 jobs.

Keep one table with the ad id, version, job id, transcript text and detected language. When legal asks which version said a given phrase, the table answers without anyone replaying clips.