MAI-Transcribe-2-Streaming partials vs Sume STT word timings

What is the difference between a streaming transcriber and a job-based one? Microsoft says MAI-Transcribe-2-Streaming returns its first hypotheses, called partials, in just over 100 ms and covers 60 languages. Sume STT 1.0 takes a public audio URL, runs as a job and always returns word-level timings.

The first suits a live caption or dictation box. The second suits a finished recording that you want to cut, subtitle or search.

Partials and finals

A streaming model emits provisional text that it later revises, then a final transcript. Microsoft reports accuracy rankings for both partial and final transcripts on Artificial Analysis, and an introductory price of $0.54 per hour of audio through the end of the year.

A job-based transcriber has no provisional stage. You submit a file and read one result when the job completes, with no text on screen while it runs.

What Sume returns

The Sume request needs a public HTTPS audio_url. language_code is optional and omitted means auto-detect. duration_seconds, from 1 to 600, only improves the usage reservation; leave it out and Sume reserves one minute. Word timings come back on every request, with no flag, and adding segmentation also returns sentence segments with a default 70 ms lead past each sentence's last word, adjustable from 0 to 500 ms.

The hourly figure in the table is plain arithmetic: $0.01 times 60 minutes. The two prices are not like for like, since one is a streaming service in preview terms and the other is a job with a catalog rate; compare them for your workload rather than as a ranking.

Choosing by the job you have

Pick on what the user sees, not on the headline price.

• Live captions, dictation or an agent that must hear a caller as they speak: a streaming model, because a job cannot show text while the speaker talks.
• Subtitles, clips and search over a finished video: a job, because timings for every word are what you need, and the delay is invisible.
• Both: stream during the session, then run the saved recording through a job for the clean timed transcript you publish.

Waiting for a job

A Sume job can be waited on for up to 30 seconds in sync mode, or polled, or delivered by webhook, as described in Sume jobs and results. For a long recording use webhook or polling rather than a held request. The cost side of the same comparison is in the cost of one hour.