MAI-Transcribe-2-Streaming is a preview with no SLA: add a fallback

MAI-Transcribe-2-Streaming is in public preview, and Microsoft's Learn page says the preview is "provided without a service-level agreement, and is not recommended for production workloads" (read 2026-10-08). If a product depends on live transcripts, the practical hedge is to record the call audio and keep a second path that can produce the transcript afterwards.

Sume can be that second path for recorded audio: its STT 1.0 route takes a public HTTPS audio URL, runs up to 600 seconds per job and costs about $0.01 per audio minute. It is not a live fallback, because Sume has no streaming transcription.

What Microsoft says

The same note covers MAI-Voice-2.1 and Flash on their Learn page. The launch post gives the streaming price as $0.54 per hour of audio "through the end of the year" and says the model ranks no. 1 on Artificial Analysis for final and partial transcripts. A ranking is not an uptime commitment, and Microsoft's pages do not state one.

Preview features "might have constrained capabilities" per the same note, so also pin your SDK version and re-test before each upgrade.

A fallback path on Sume

Save the raw call audio to storage you control and expose it at a public HTTPS URL for each 10-minute slice. Then submit a Sume STT job per slice. A failed live stream costs you the live captions, not the record of the call.

Send an Idempotency-Key on every submit so a client retry returns the same job instead of billing twice. The STT route returns text and word timings (word, start, end in seconds), so you can rebuild a transcript, search it and cut clips from it.

import os, requests

H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}",
     "Idempotency-Key": "call-8841-slice-0"}
r = requests.post(
    "https://api.sume.com/v1/stt-1.0/transcribe",
    headers=H,
    json={"audio_url": "https://media.example.com/call-8841-0.m4a",
          "duration_seconds": 600, "mode": "async"},
)
r.raise_for_status()
print(r.json()["data"]["request_id"])

Limits to state up front

Sume STT 1.0 returns no speaker labels, and has no phrase list or style option. A call-center review that needs agent versus customer separation must record the two sides as separate files, or label speakers some other way.

Cost for the fallback is small: a 30-minute call is three 10-minute jobs at 10 cents each, 30 cents, billed only if you run them.

A simple runbook

Define what failure means before it happens. If the live stream drops, show the agent a banner and keep recording locally. If the stream returns errors for more than a set number of seconds, stop trusting live text and mark the call for after-the-fact transcription. When the call ends, enqueue the saved audio slices and attach the transcript to the record within the hour.

Review the preview terms on the Microsoft page each time you renew a roadmap. Moving from preview to general availability usually changes the terms, the price and sometimes the model name, and a fallback that you have already tested makes that migration a planned change rather than an incident.