Introducing MAI-Voice-2: The Most Expressive Text-to-Speech Model to Date

Research context and background

Microsoft has launched MAI-Voice-2, a new text-to-speech model designed to create the most natural and expressive human-like voices available today. This tool builds on the success of its predecessor, MAI-Voice-1, by improving sound quality, expanding language support, and adding deeper emotional control. It is built with safety features to ensure that voices are used responsibly and only with proper consent.

Benefits

MAI-Voice-2 offers several key advantages for developers and creators. First, it supports 15 different languages, including English, Spanish, French, German, Hindi, and Chinese, while maintaining high naturalness in each. Users can control the emotional tone of the voice using simple tags like sad, whispered, or excited. The model also allows developers to create custom voices using just 5 to 60 seconds of reference audio without needing to retrain the system. In testing, this new version was preferred over the older model 72% of the time. It also keeps the speaker identity consistent throughout long recordings, which is perfect for audiobooks or lectures. Additionally, it can switch between languages smoothly within a single sentence for pairs like Hindi-English or Spanish-English.

Use Cases

This technology is designed for many real-world applications. Companies can use it to give branded voices to customer support agents, virtual assistants, and mobile apps. Creators can use it to generate characters for video games, podcasts, and virtual reality experiences. It is also a powerful tool for accessibility, helping visually impaired users listen to content or assisting people with speech impairments. Educators can use the model to create virtual instructors for courses and simulations. Finally, individual creators can turn written text into audio using their own voice without needing a professional recording studio.

Pricing

Pricing details for MAI-Voice-2 are not publicly available in the provided information. The model is currently accessible through Microsoft Foundry and the Microsoft Foundry API documentation.

Vibes

Public reception highlights the model as a significant leap forward in voice technology. Early preference tests show that users strongly favor MAI-Voice-2 over previous versions. The team behind the project, the Superintelligence Team, describes the model as the most expressive and natural-sounding text-to-speech system built to date. The inclusion of robust consent guardrails has been noted as a positive step toward trustworthy and responsible deployment.

Additional Information

MAI-Voice-2 is available in Microsoft Foundry and is being integrated into Visual Studio Code and the Dynamics 365 Contact Center. The model was developed by the Superintelligence Team, a group focused on creating high-impact AI solutions. The team has an ambitious roadmap and is working with next-generation hardware to support their mission of helping billions of users. Developers can try the model through the DuoAI experimental experience or by accessing the API documentation and Cookbook.