MAI-Transcribe-2 vs Sume STT 1.0: keywords, diarization, price

MAI-Transcribe-2 has more request-time controls than Sume STT 1.0 and a lower limited-time price. Sume's STT is $0.01 per audio minute, or $0.60 an hour, with word timings always returned and no keyword-biasing field.

Side by side

Microsoft's model page and announcement describe 60 languages, a 5.2 percent average word error rate on FLEURS, diarization, word-level timestamps, keyword biasing, verbatim and clean styles, code-switching and automatic language identification. Its price is $0.10 per hour as a limited-time rate through the end of 2026. Sume's STT 1.0 page and API reference give the rest.

A Sume request

Send an audio_url, an optional language_code and, if you know it, duration_seconds. The response is the usual async job.

curl -sS -X POST https://api.sume.com/v1/stt-1.0/transcribe \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: call-0001" \
  -d '{"audio_url":"https://example.com/call.mp3","language_code":"en","duration_seconds":240}' \
  | jq '.data | {request_id, status_url, result_url}'

How to choose

If your audio is full of product names that a general model misspells, keyword biasing is the feature that decides, and Sume does not offer it as a field. If you need word timings for captions and a hosted job API alongside TTS and rendering, Sume covers it in one account.

Limits

The Microsoft figures are vendor claims from its own pages and I did not run an accuracy test. The $0.10 per hour rate ends in 2026; ask what follows.