Microsoft MAI voice models land on Vercel AI Gateway, callable through one API with no inference markup
Synopsis
Vercel and Microsoft AI (MAI) have partnered to bring MAI models to AI Gateway, making the platform one of a limited number of places that offer access to MAI models. Three audio models are supported: MAI-Voice-2.1 generates expressive speech in 23 languages and keeps a consistent speaker across longer passages, MAI-Voice-2.1-Flash targets lower-latency interactions, and MAI-Transcribe-2 Streaming returns partial transcripts as audio arrives. AI Gateway bills these models at their listed rates, with no platform fee or markup on inference.
Interpretation
Developers can now call Microsoft AI's audio models through Vercel's AI Gateway instead of integrating separately with MAI's model service. AI Gateway did not previously offer MAI models; with the integration, the platform becomes one of a limited number of platforms offering access to them. The models available are MAI-Voice-2.1, MAI-Voice-2.1-Flash and MAI-Transcribe-2 Streaming, covering long-form speech generation, low-latency speech generation and streaming transcription respectively.
The three models cover two task types: two speech-synthesis models turn text into speech, and one transcription model turns audio into text. MAI-Voice-2.1 generates speech in 23 languages and keeps a consistent speaker across longer passages, while the Flash variant brings the same multilingual speech generation to lower-latency interactions. MAI-Transcribe-2 Streaming returns partial transcripts while audio is still arriving, so applications can show live captions or follow a conversation without waiting for the full recording.
Billing follows the models' listed rates, and AI Gateway adds no platform fee or markup on inference. AI Gateway itself provides one API for calling models, tracking usage and cost, and viewing request traces, and it supports routing, retries and failover across available providers. The integration example uses generateSpeech in AI SDK 7 to call microsoft/mai-voice-2.1-flash, selecting Microsoft's Harper voice through the voice parameter.
Perspective
The report is limited to what the Vercel and Microsoft AI partnership announcement states: three MAI audio models are callable on AI Gateway, billed at listed rates with no markup, with call examples for speech generation and streaming transcription.
The announcement gives no quantitative measures of speech naturalness, latency or transcription accuracy for the three models, and does not state the range of other platforms where MAI models are available; partial transcripts can change as more audio arrives, so applications must replace displayed text when a new partial arrives.