90 voice notes of 2 minutes: transcript cost, MAI vs Sume STT

Transcribing 90 voice notes of 2 minutes each is 180 audio minutes, or 3 hours, which costs $1.62 on MAI-Transcribe-2-Streaming and $1.80 on Sume STT 1.0.

Voice notes are short and varied, so the main settings are the language hint and the duration hint.

The arithmetic

MAI-Transcribe-2-Streaming is priced per hour, Sume STT 1.0 per audio minute, so the sums convert minutes to hours first. The Sume rate is the public rate on the Sume docs; confirm the live number in GET /v1/catalog.

Per minute, MAI works out to $0.009 ($0.54 / 60) against Sume's $0.01, so Sume is about 11 percent higher. On this run the difference is $0.18. Microsoft's rate is an introductory price that its page says runs through the end of the year, and the page does not say what the price becomes afterward, so a budget past December should not assume it.

Mixed languages

If a note mixes languages, omit language_code and let auto-detect run; a wrong hint can hurt more than none.

Send duration_seconds: 120 on each job; without it Sume reserves one minute, which is half of what a 2-minute note needs.

• Keep the note id as the idempotency key.
• Store the words array, not just the text, if you want to search by time later.

Limits that shape a Sume run

Sume does not list MAI-Transcribe-2-Streaming, so the MAI row is a price reference. Sume STT 1.0 takes a public HTTPS audio_url and returns the transcript with word timings; there is no flag for timings because they always come back. Sentence segments are available when you send segmentation.

• duration_seconds is an optional hint from 1 to 600; leave it out and Sume reserves one minute, so send it for a long file.
• language_code is a hint such as en or ko; leave it out for auto-detect.
• diarize and tag_audio_events are fixed server-side, so do not send them; there are no speaker-label controls.
• The MAI price is an introductory rate that Microsoft says runs through the end of the year.

Running it

Submit with stt_create and an idempotency_key, wait with jobs_wait, then read jobs_result for the text and words. Cut recordings longer than 10 minutes at a pause first; audio detach pulls the track out of a video and Timeline audio split slices a Sume-hosted file into ranges.

The figures are list arithmetic as of 2026-10-08. A one-time check of your own audio is worth more than a benchmark, so run one representative file through each option before you commit a budget.