Microsoft AI: MAI-Voice-2

MAI-Voice-2 is an expressive text-to-speech model from Microsoft AI, released by the provider on 2 June 2026. It sits in the speech category, so its job is turning written text into spoken audio rather than generating or editing text itself. If you are building anything that needs a synthetic voice, this is the kind of model you would evaluate alongside other text-to-speech options.

The model takes text as its only input and produces speech as its only output. It is described as expressive, which matters for applications where flat, robotic delivery would hurt the experience. Microsoft positions it for conversational assistants, media narration, accessibility, education, and other long-form voice applications, and it supports 15 languages. No context length is listed for this model, so treat token-based context planning as not applicable here. The available voices are en-US-Harper, es-MX-Valeria, fr-FR-Soleil, and de-DE-Klaus, each tied to the MAI-Voice-2 model.

Those capabilities point to a fairly clear set of use cases. Conversational assistants can use it to speak replies back to users in a natural voice. Media and content teams can use it for narration over video, articles, or other long-form material. Accessibility tools can convert written content into audio for users who prefer listening. Education products can read lessons, exercises, or study material aloud. Because the model handles long-form voice work, it fits projects where you need sustained speech output rather than a single short clip.

You call it through api14's OpenAI-compatible API using the model id "microsoft/mai-voice-2". Because the interface follows the OpenAI format, you can point existing OpenAI-style client code at api14 and swap in this model id. The model lists no supported parameters, so do not assume tuning knobs such as temperature or top_p will have an effect. Voice selection is the main control surface, and you choose from the four voices listed above. Note that maxOutput is reported as 0, so plan your requests around the audio you actually need rather than relying on a documented output ceiling.

If your task is text generation, translation, classification, or anything that returns written output, this model is the wrong fit, since it only emits speech. It is also a poor choice if you need a voice outside the four listed or a language beyond the 15 it supports. For those cases, look at a text model or a different speech model on api14.