MAI-Voice-1

What MAI-Voice-1 is

MAI-Voice-1 is a speech generation model from Microsoft AI, announced on 28 August 2025 as the first model in the MAI family ^[1]. Microsoft described it as delivering high-fidelity, expressive audio in both single-speaker and multi-speaker scenarios ^[1]. It is the model behind several of Microsoft's spoken-audio products, from daily briefings to podcast generation ^[1][2].

Speech generation of this kind turns text into audio that has to do more than pronounce words correctly: it has to carry tone, pacing and a consistent voice across a long piece of writing. Those two requirements — expressiveness and consistency — run through most of what Microsoft says about this model ^[1][2].

Speed as the selling point

The headline claim at announcement was throughput. Microsoft described the model as "lightning-fast", saying it can generate a minute of audio in under a second on a single GPU ^[1]. When the model reached developers in Microsoft Foundry, the company quoted a similar figure: about 60 seconds of audio in roughly one second ^[2].

Speed of that order changes what products are practical. A podcast episode or a narrated article is minutes long, and generating it in seconds rather than minutes makes on-demand audio features feel responsive to the person waiting for them.

Where it was used at launch

At launch MAI-Voice-1 powered Copilot Daily and Copilot Podcasts, and Microsoft offered it in Copilot Labs so people could try expressive speech and storytelling demonstrations ^[1]. The demonstrations listed at the time included a choose-your-own-adventure story and a bespoke guided meditation for sleep ^[1]. Both are cases where the voice carries the experience: a story that branches needs distinct, recognisable speakers, and a meditation depends on a calm, steady delivery.

Release to developers

On 2 April 2026 the model was released to developers in Microsoft Foundry and the MAI Playground ^[2]. That release came alongside two other models, MAI-Transcribe-1 and MAI-Image-2 ^[2].

In Foundry the model supports custom voices built from just a few seconds of audio, keeps a speaker's identity intact across long-form content, and generates 60 seconds of audio in about one second ^[2]. Microsoft positioned it as its top-tier voice generation model, built specifically to preserve speaker identity even across long-form material ^[2]. Holding a voice steady over many minutes is the harder half of that job: a short clip can sound convincing while a long one drifts in pitch, rhythm or timbre.

Pricing at the Foundry launch started at $22 per million characters ^[2][3].

Position in the MAI line

MAI-Voice-1 was the first of the MAI models to be announced, and Microsoft continued to use it in Copilot Audio Expressions and Copilot Podcasts after the developer release ^[1][2]. It was followed in June 2026 by MAI-Voice-2, which covers 15 languages ^[2]. Read together, the two releases show the model moving from a launch-day demonstration inside Copilot to a priced, developer-facing service with custom voice creation, and then to a successor with wider language coverage.