Details
- Microsoft AI is launching MAI-Transcribe-2-Streaming, a real-time speech-to-text model designed to transcribe audio continuously as it arrives.
- Microsoft says the model is 55% faster and 60% cheaper than competing models in its class, with Mustafa Suleyman’s post specifically comparing it with ElevenLabs.
- The model generates initial transcript hypotheses within the low hundreds of milliseconds, then refines them as more speech context becomes available.
- MAI-Transcribe-2-Streaming supports 60 languages and automatic language detection, targeting live applications such as captions, call transcription, accessibility, and voice-enabled agents.
- The models are available immediately in Microsoft Foundry, Microsoft’s platform for discovering and deploying AI models.
- Microsoft is also making the models available through developer platforms including Vercel and OpenRouter, expanding access beyond Azure-based deployments.
- The announcement positions the transcription model as infrastructure for agents that need fast, continuously updated speech input rather than batch transcription after a recording ends.
Impact
The launch strengthens Microsoft’s position in the developer infrastructure market by combining low-latency transcription with distribution through Foundry, Vercel, and OpenRouter. The claimed cost and speed improvements could lower the operating cost of voice agents and increase adoption in call, accessibility, and meeting applications. Competition remains active: Microsoft Foundry already offers real-time transcription models from OpenAI, while ElevenLabs and other speech providers continue to compete on voice and audio tooling. The comparison is therefore best understood as a price-and-latency challenge, not proof of a categorical market lead.
