Microsoft's new speech model beats Whisper at bargain prices

Speech-to-text models work great in sterile demo videos. But throw in a whirring cafe espresso machine, overlapping speakers in a messy meeting, or a spotty phone connection, and transcripts fall apart fast.

Microsoft just launched MAI-Transcribe-1, a new multilingual speech-to-text model now available in public preview on Microsoft Foundry and the Microsoft AI Playground. It is designed specifically to handle noisy, chaotic, real-world audio across 25 languages.

Beating Whisper and cutting latency

According to Microsoft, MAI-Transcribe-1 achieved the lowest Word Error Rate on the 25-language FLEURS benchmark, beating out Scribe v2, Whisper-large-V3, GPT-Transcribe, and Gemini 3.1 Flash-Lite.

A footnote from Microsoft clarifies that across the top 25 languages by product usage, MAI-Transcribe-1 took first place in 11 core languages. On the remaining 14, it beat Whisper-large-v3 across the board and outperformed Gemini 3.1 Flash on 11 of them.

Raw accuracy aside, Microsoft focused heavily on pipeline efficiency. The model delivers batch transcription speeds 2.5x faster than Azure's existing Fast offering, while keeping latency low enough to handle real-time dictation, live captioning, and conversational voice agents.

It is priced at $0.36 per hour of audio, which Microsoft touts as the best price-to-performance ratio among major hyperscalers.

Taking over Teams and Copilot

Microsoft highlighted challenging edge cases in its launch examples: changing a flight in a crowded cafe, dealing with code-switching between Spanish and English in an office, and transcribing requests at a loud concert. The model handled background chatter, low bitrates, and cross-talk without losing context.

MAI-Transcribe-1 is already in phased rollouts inside Microsoft Teams and Copilot's Voice mode to power downstream meeting recaps and action items.

It is launching alongside two other in-house models: MAI-Voice-1 for speech synthesis ($22 per 1M characters) and MAI-Image-2. Microsoft pitches the combination of MAI-Transcribe-1 and MAI-Voice-1 as a full-stack audio layer for custom voice agents.

Why it matters

Accurate audio ingestion is the silent make-or-break bottleneck for voice agents. If an underlying model mishears an intent because of cafe background chatter, the most sophisticated LLM in the world cannot save the user experience.

For developers building call-center analytics, meeting summarizers, or real-time voice bots, $0.36 an hour combined with 2.5x batch speeds makes this an immediate no-brainer to benchmark against Whisper API setups. Real-world resilience against cross-talk is notoriously hard to train, and this directly targets that pain point.

Strategically, this marks another clear step in Microsoft building out its own proprietary MAI stack rather than solely relying on OpenAI endpoints. By pushing MAI-Transcribe-1 directly into Teams and Copilot, Microsoft gets an unmatched real-world feedback loop from hundreds of millions of daily enterprise interactions. The limitation remains its initial 25-language scope, but for mainstream global workloads, Microsoft just put serious price and speed pressure on both OpenAI and Google Cloud.