30 customer calls of 8 minutes: transcription bill, MAI vs Sume

Thirty recorded calls of 8 minutes each come to 14,400 seconds, or 4 hours of audio. That costs $2.16 on MAI-Transcribe-2-Streaming at its launch price and $2.40 on Sume STT at 1 cent a minute. The 24 cent gap is not a reason to choose either.

The bill

The deciding question is whether the calls are already recorded or still live. A recorded call is a file, and Sume STT is built for files.

One call, one job

An 8-minute call is 480 seconds, under the 600-second cap of one Sume STT job, so each call is one job with no splitting. Post the call's public HTTPS URL as audio_url, add duration_seconds: 480 so the usage reservation matches the call, and read the text and word times when the job completes. Without a duration Sume reserves one minute.

Thirty jobs go out with 30 different Idempotency-Key values. A key made from the call id is a good choice, because a retried submit then returns the same job instead of billing the call again.

Search inside a call

Ask for sentence segments with segmentation: {"mode": "sentence"} and each sentence arrives with a start and end time. A support team can then jump from a flagged sentence straight to that moment in the recording. The segments are time ranges over your file, and Sume produces no cut audio files for them.

Word timings are always returned. The speaker split and audio-event tagging are fixed on the server, and a request that sends diarize or tag_audio_events is rejected, so do not plan a pipeline around changing them.

Check the first five calls by reading the transcripts next to the audio. Telephone audio with crosstalk and hold music is the hard case for any recognizer, and a short read tells you whether to trust the batch before you pay for all thirty.

A call over 10 minutes needs a split before you transcribe. A 14-minute call is 840 seconds, which is two jobs of 420 seconds, and the stitch step must add the second job's start time to its word times.

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.

Live or recorded

Streaming transcription is the better match for live calls, where text must appear while the caller speaks. If your goal is a live agent assist panel, that is Microsoft's streaming model's job and Sume STT is not a drop-in for it. If your goal is a nightly batch that tags calls for a quality review, a file job is simpler, because you never hold a connection open.

Calls are personal data. Put the recordings behind URLs you control, make the links short-lived, and keep transcripts under the same retention rule as the audio. That is a policy matter on your side, whichever service reads the file.

Finally, store the job id with each call record. If a customer asks what was said on a call, the job result holds the text and word times, and you can re-read it without paying to transcribe the call again.