Two-hour council meeting transcript: 12 Sume jobs, $1.20

A two-hour public meeting costs $1.20 to transcribe on Sume: 7,200 seconds is twelve 600-second STT 1.0 jobs at 10 cents each. The same two hours at Microsoft's listed MAI-Transcribe-2 batch rate of $0.10 per hour is $0.20, and at the MAI-Transcribe-2-Streaming rate of $0.54 per hour is $1.08 (all read 2026-10-08, both Microsoft prices limited-time through the end of the year).

Sume is not the cheapest line in that table. It is the one you can call with a single HTTPS request per slice and no cloud subscription.

Cost side by side

Microsoft's post describes the $0.10 rate as "a limited-time offer until the end of the year" and the streaming rate as an introductory price through the same date. Neither page states a price for 2027, so do not build a multi-year budget on them.

Sume's rate is flat: about $0.01 per audio minute, rounded up per job, with a 10-cent cap per 600-second job. A meeting that is exactly 7,200 seconds cleanly fills twelve jobs.

Cutting the recording

Sume STT reads a public HTTPS audio URL and accepts at most 600 seconds. The audio-detach route will not take a source over 1,800 seconds, so for a two-hour file cut the slices yourself. This ffmpeg command writes 16 kHz mono WAV slices of 600 seconds, the shape the audio-detach docs recommend for speech-to-text. Upload each slice, then submit one STT job per URL.

ffmpeg -i meeting.mp4 -vn -ac 1 -ar 16000 -c:a pcm_s16le \
  -f segment -segment_time 600 -reset_timestamps 1 slice_%02d.wav
ls slice_*.wav | wc -l   # expect 12 for a 7,200 s recording

What a clerk still has to do

The STT result has text and word-level timings, with no speaker labels. For meeting minutes you will need to mark who spoke, either by listening or from the agenda order. A transcript with timestamps lets you link each agenda item to the second it starts.

Run slices in parallel; each is its own job. Offset each slice's word times by its start (slice index times 600) to get a single timeline, the way the 45-minute talk post does.

Publishing the record

A transcript posted by a public body is a record, so label it as machine-generated and unreviewed until staff have checked it. Keep the audio and the transcript together under the same date and agenda number. Because Sume's word timings are in seconds from each slice's start, store the slice index alongside every timestamp or normalize them to one timeline before you save.

For long-term budgeting, remember the Microsoft prices in the table are limited-time. A city that transcribes 50 meetings a year of two hours each spends $60 on Sume at the published rate; the point of the comparison is that the transcription bill is small either way and the staff time to review it is not.