MAI-Transcribe-1.5 is Microsoft AI's in-house speech-to-text model, built for production transcription across 43 languages with automatic language detection and domain-specific "entity biasing." It posts a 2.4% Word Error Rate on the Artificial Analysis leaderboard, ranks #1 on the FLEURS multilingual benchmark, transcribes an hour of audio in under 15 seconds, and costs $0.36 per hour of audio through Microsoft Foundry.
What Is MAI-Transcribe-1.5?
MAI-Transcribe-1.5 is the second-generation speech-to-text model built by Microsoft's in-house MAI (Microsoft AI) team. It's designed to turn noisy, real-world audio — meetings, calls, dictation — into accurate, domain-aware transcripts, and it's positioned explicitly for production workflows rather than as a research demo. The model expands language coverage from 25 to 43 languages compared to the original MAI-Transcribe-1, adding 18 new languages (including a cluster of South Asian and European languages) without giving up accuracy, according to Microsoft.
It's available now through Microsoft Foundry (Azure AI Foundry) as a generally-available model, through the MAI Playground for hands-on testing, and via OpenRouter. Microsoft has also folded it into first-party products, including Copilot, Teams, GitHub, and Dynamics 365 Contact Center, where it handles captioning and call transcription behind the scenes.
Benchmark Results: FLEURS and Artificial Analysis
Microsoft evaluates the model on two independent benchmarks it doesn't control, which is worth noting since most model-launch benchmarks are self-reported without third-party context:
| Benchmark | Result | Notes |
|---|---|---|
| FLEURS (multilingual) | Best-in-class WER across 43 languages; #1 ranked | Beats Scribe v2 in 29 of 43 languages, Gemini 3.1 Flash in 34 of 43, and GPT-Transcribe in 27 of 40 |
| Artificial Analysis | 2.4% WER, #3 overall rank | Described by Microsoft as the leader when accuracy is weighed against speed |
| Long-audio speed | 1 hour of audio transcribed in under 15 seconds | Up to 5x faster than Gemini 3.1, Scribe v2, and GPT-4o-Transcribe on long-form audio; 5.7x faster than MAI-Transcribe-1 |
The speed numbers matter as much as the accuracy numbers here. A model that's marginally more accurate but takes several times longer to process an hour-long call recording is a worse fit for a production call-analysis pipeline than one that trades a fraction of a percentage point of WER for a 5x throughput gain.
Entity Biasing: The Headline Feature Over MAI-Transcribe-1
The main capability MAI-Transcribe-1.5 adds over its predecessor is entity biasing (also called keyword or contextual biasing). Developers can prime the model with up to 200 domain-specific terms — meeting attendee names, product names, medical vocabulary, internal acronyms, customer-specific terminology — and the model will bias its predictions toward those entries whenever the acoustic signal is ambiguous.
Microsoft reports this produces up to a 30% WER reduction on the FLEURS benchmark when biasing is enabled. Notably, the model doesn't blindly force a match every time a biased term is nearby in the audio; it uses surrounding context to decide when biasing should actually apply, which is meant to avoid the common failure mode of keyword-biasing systems overcorrecting and inserting a primed term where it doesn't belong.
This is the feature that makes the model interesting for regulated or jargon-heavy domains specifically called out by Microsoft: pharmaceutical dictation, medical documentation, and customer service, where a single misheard drug name or product SKU can matter far more than a generic WER percentage suggests.
Pricing and Availability
MAI-Transcribe-1.5 is priced at $0.36 per hour of transcribed audio through Microsoft Foundry (Azure AI Foundry), positioned as a low-cost, high-throughput option relative to comparable transcription APIs. You can test it without committing to an integration through the MAI Playground at playground.microsoft.ai, or call it directly via Microsoft Foundry or OpenRouter for production use.
How It Fits Into Microsoft's Broader MAI Push
MAI-Transcribe-1.5 didn't ship alone — it's part of a wider rollout of Microsoft-built MAI models across text, image, voice, and speech that Microsoft has been shipping into Foundry. The pattern is consistent with Microsoft's broader strategy of building first-party models for cost- and latency-sensitive workloads (transcription, voice, smaller coding models) while continuing to rely on partner frontier models for the hardest reasoning tasks. If you're evaluating transcription options, it's also worth comparing this against Google's competing release in the same space.
FAQ
What is MAI-Transcribe-1.5?
MAI-Transcribe-1.5 is Microsoft AI's speech-to-text model, built in-house for production transcription workloads. It supports 43 languages, automatic language detection, and domain-specific entity biasing, and is available through Microsoft Foundry, the MAI Playground, and OpenRouter.
How accurate is MAI-Transcribe-1.5?
It posts a 2.4% Word Error Rate on the Artificial Analysis leaderboard and ranks #1 on the FLEURS multilingual benchmark across 43 languages, beating Scribe v2, Gemini 3.1 Flash, and GPT-Transcribe in the majority of language-by-language comparisons on FLEURS.
How much does MAI-Transcribe-1.5 cost?
It's priced at $0.36 per hour of audio through Microsoft Foundry.
What languages does MAI-Transcribe-1.5 support?
43 languages, up from 25 in MAI-Transcribe-1. Microsoft added 18 new languages in this release, including a number of South Asian and European languages, while keeping accuracy consistent across the expanded set.
What is entity biasing in MAI-Transcribe-1.5?
Entity biasing lets developers supply up to 200 domain-specific terms — names, product terms, medical vocabulary — that the model will lean toward recognizing when the audio is ambiguous. Microsoft reports up to a 30% WER reduction on FLEURS when it's enabled, and the model uses context to avoid forcing false matches.
How fast is MAI-Transcribe-1.5 compared to other transcription models?
It can transcribe an hour of audio in under 15 seconds and is up to 5x faster than Gemini 3.1, Scribe v2, and GPT-4o-Transcribe on long-form audio, and 5.7x faster than its predecessor, MAI-Transcribe-1.
Where can I try MAI-Transcribe-1.5?
You can test it directly at the MAI Playground (playground.microsoft.ai), or integrate it via Microsoft Foundry or OpenRouter. It's also already built into Copilot, Teams, GitHub, and Dynamics 365 Contact Center.
Is MAI-Transcribe-1.5 meant for regulated industries like healthcare?
Microsoft specifically calls out pharmaceutical and medical dictation as target use cases, leaning on the entity-biasing feature to correctly capture drug names and clinical terminology that generic transcription models tend to mishear.
Further reading:
- Gemini 3.5 Transcribe Explained: Google's New Speech-to-Text Model for Voice Apps (2026)
- What Is RVC (Retrieval-based Voice Conversion) and How to Run It 100% Locally (2026 Guide)
- Top AI Models to Use in Cursor IDE for App Development (2026 Guide)
- Multi-Agent Orchestration and Governance: Why Most AI Agent Pilots Never Reach Production