Microsoft adds streaming transcription and two text-to-speech models to Azure AI Foundry

Microsoft has added three speech models to its MAI lineup in Azure AI Foundry, following the non-streaming MAI-Transcribe-2 released earlier this month. One model transcribes live speech as it arrives; two convert text to spoken audio with different quality and cost trade-offs.

MAI-Transcribe-2-Streaming handles 60 languages with automatic language detection. It returns partial transcripts before a speaker finishes. Microsoft says the first hypotheses arrive within a few hundred milliseconds of receiving audio, then refine as more context arrives and commit a stable version once the utterance ends. Microsoft claims the model ranks #1 for accuracy on both partial and final transcripts on the Artificial Analysis leaderboard, with words appearing as early as 320 milliseconds after being spoken versus more than 500 milliseconds for the closest competitor. Those figures come from Microsoft's own reporting of a third-party benchmark, so vendor-reporting caveats apply. Microsoft frames the practical difference as whether an application can act on a request mid-sentence — identifying a caller's issue, looking up a record, preparing a tool call — rather than waiting for a complete transcript.

The text-to-speech side includes two models. MAI-Voice-2.1 supports 23 languages with a consistent voice identity across them: the same voice adapts pronunciation and delivery per language rather than carrying one accent everywhere. MAI-Voice-2.1-Flash is the same model family optimized for latency and high volume. Microsoft says Flash is 55% faster and 60% less expensive than competing models in its class, with the comparison sourced from ElevenLabs' documentation — again, a vendor-asserted figure rather than an independent evaluation.

Pricing: MAI-Voice-2.1 runs $22 per 1M characters; Flash runs $15 per 1M characters. MAI-Transcribe-2-Streaming is priced at an introductory $0.54 per hour of audio through the end of the year. The earlier MAI-Transcribe-2 remains available for workloads needing speaker diarization and word-level timestamps, which the streaming model apparently does not offer.

All three models are accessible through Microsoft Foundry and Azure Speech, and through OpenRouter, Vercel, and LiveKit.

The announcement does not answer two questions: whether the streaming model's accuracy holds up outside benchmark conditions, and whether $0.54 per hour of audio is a long-term rate or a promotional one. Transcription pricing is the line item most likely to balloon for a voice-agent deployment running constantly, so that answer matters more to production users than any leaderboard position.