Microsoft recently unveiled its first in-house AI models: MAI-Voice-1 AI and MAI-1-preview. The MAI-Voice-1 model is designed to produce one minute of audio in less than one second using a single GPU. The MAI-1-preview aims to provide insights into future offerings for the company’s Copilot assistance platform.
Users can test MAI-Voice-1 via Copilot Labs, where they can input text for the AI to generate speech, along with options to modify voice and speaking style. The MAI-1-preview has been trained on approximately 15,000 Nvidia H100 GPUs and is intended to assist users with responding to everyday inquiries by following instructions effectively.
Microsoft’s AI chief, Mustafa Suleyman, mentioned in a previous podcast episode that the company’s internal AI models focus on consumer use rather than enterprise applications. He highlighted the importance of optimizing the models for consumer needs, utilizing extensive data for better performance.
The rollout of MAI-1-preview is planned for specific text use cases within the Copilot AI assistant, which currently utilizes OpenAI’s large language models. Additionally, Microsoft has begun public testing of the MAI-1-preview model on the AI benchmarking platform LMArena.
In its announcement, Microsoft expressed aspirations for future developments, emphasizing the potential for specialized models to address diverse user intents and use cases, which they believe will unlock significant value in AI applications.
Source: https://www.theverge.com/news/767809/microsoft-in-house-ai-models-launch-openai