AI & Machine Learning
·By Seedwire Editorial·

Microsoft MAI-Transcribe-2 launch price runs through year-end

Microsoft's September 3 MAI-Transcribe-2 announcement lists a launch price of $0.10 per hour of audio through the end of 2026. The version checked on September 19 specifies that deadline, but does not give the rate that follows. Launch coverage had left the offer's end date unresolved.

The model supports 60 languages, speaker diarization, word-level timestamps, vocabulary biasing and configurable transcript styles. Microsoft reports a 5.2% average word error rate on FLEURS across 60 languages and second place on the Artificial Analysis word error rate leaderboard. These are attributed benchmark claims, not a guarantee for a customer's recordings.

Separate recognition from speaker accuracy

A transcription evaluation needs more than a single score. Check the words against a reference transcript, then separately check which speaker receives each passage. A recording can contain the right words while assigning an instruction or commitment to the wrong participant. That distinction matters if the output feeds meeting notes or a customer record.

Use recordings representative of the intended workload, with permission to process them, and keep a separate test set when adjusting vocabulary hints. Repeatedly tuning against the same examples can hide failures on new recordings. Review each required language separately so that a strong result in one does not conceal a weak result in another.

Budget for usable transcripts

The launch rate is a useful input to a budget, but it cannot establish the cost of a finished workflow. Track how much audio must be reprocessed and how much correction the output needs. A lower API bill may not compensate for additional review time.

Before extending a cost forecast beyond the offer, confirm the applicable rate with the provider. Before replacing a working service, compare both systems on the same recordings and acceptance criteria. The useful outcome is a transcript that preserves the words and speaker assignments the application needs, with a cost model that survives the introductory offer.

Correction, September 19, 2026: This article was updated against the current announcement, and unrelated article links were removed.

MAI-Transcribe-2
Microsoft
speech recognition
Seedwire Newsletter

Follow Seedwire by email

Request Seedwire news emails. There is no guaranteed delivery schedule. You can withdraw your request through the privacy contact.

By selecting Subscribe, you request Seedwire news emails. Privacy and withdrawal.