MAI-Voice-2.1 and MAI-Transcribe-2: what Sume lists for voice and STT

Microsoft AI's news page names MAI-Voice-2.1 and MAI-Transcribe-2 (read 2026-10-08). Sume does not list either one, so you cannot call them through Sume. This post gives no Microsoft prices or specifications; it states what Sume does list for text-to-speech and speech-to-text so you can compare against any quote you get elsewhere.

What Sume lists for voice

The Sume TTS router lists Sonic ids: sonic-3.6, sonic-3.5, sonic-3, the alias sonic-latest (for sonic-3.6), and the beta sonic-preview. Each has a 20,000-character request cap and a list price of 38 micro-dollars per character. Sume bills the list price times 1.25, rounded up to a whole cent for each job. Speech-to-text is Sume STT 1.0 at $0.01 per audio minute.

How to compare without guessing

Do not compare a per-character price to a per-hour price by eye. Convert both to your own workload. For narration, use your average script length. A 1,000-character script on Sume is 47,500 micro-dollars, billed as 5 cents. A 20,000-character script is $0.95. For transcripts, a 60-minute recording is 60 x $0.01 = $0.60.

Then ask what the quote leaves out. On Sume, STT through video inspect also bills compute and a platform fee, TTS rounds up to a cent per job, and Timeline renders are $0.10 per output minute. Put each vendor's total next to yours for the same 10 episodes.

What would change this post

If Sume adds either model to its catalog, the GET /v1/tts-router/models response or the Sume docs will list it, and the price will come from there. Until then, treat any claim that Sume offers a Microsoft voice or transcription model as wrong, and read the catalog yourself.

Questions to put to any voice quote

Ask what the unit is (character, second, minute), whether there is a minimum charge per request, what the request size limit is, how retakes are billed, and which model id is recorded with each output. On Sume the answers are: character, one cent per job through rounding, 20,000 characters per request, a full price per retake, and model_id on the job record.

Then run the same ten scripts through each option and compare totals, not rates.