Microsoft's latest in-house AI models show a clear strategic shift, but real-world tests reveal a persistent performance gap with established players.
Microsoft has launched three proprietary artificial intelligence models, a move seen by industry observers as a significant step toward reducing its dependence on partner OpenAI. The new models—MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2—cover speech transcription, voice generation, and image creation and are now commercially available through the Microsoft Foundry platform.
"This move marks Microsoft's effort to build its own AI technology stack," one media report commented, reflecting a view that the company is diversifying its AI capabilities. The Foundry platform now offers Microsoft's MAI series alongside models from OpenAI and Anthropic, providing customers access to multiple providers through a single API.
The company's official benchmarks claim significant performance gains. MAI-Transcribe-1 is reportedly 2.5 times faster than the existing Azure Fast product, MAI-Voice-1 can generate 60 seconds of audio in just one second, and MAI-Image-2 offers at least a twofold improvement in image generation speed. Pricing has been set at $0.36 per hour for transcription, $22 per million characters for voice generation, and starts at $5 per million tokens for image text-prompts.
For investors, the launch raises a critical question: can Microsoft's internal development close the performance gap with leading models from partners like OpenAI and competitors like Google? While Microsoft's OpenAI contract extends to 2032, the economic viability of its in-house strategy hinges on achieving competitive performance, a factor that will determine long-term returns on its significant R&D investment.
In tests, the MAI-Transcribe-1 model showed mixed results. While it accurately transcribed a scene from the film Infernal Affairs at normal speed, it failed when the audio was played at double speed. The model misinterpreted a line about "police academy" (警校) and "undercover agents" (卧底) as being about "Cambridge" (剑桥) and "accountants" (会计), completely altering the context.
The model's stability was further challenged with a more intense, fast-paced argument from the movie Cold War, where it failed to produce any output. These tests show that while the model is competent for standard speech, its performance declines with complex audio involving high speed or strong emotion, exposing a gap compared to market leaders like OpenAI's Whisper.
The other models demonstrated both strengths and weaknesses. MAI-Voice-1 produced impressively distinct audio styles, including a Shakespearean English accent with theatrical pacing and a bright, modern American accent. The model's output included fine details such as the sound of saliva, adding a high degree of realism.
MAI-Image-2, which ranks third on the Arena.ai user leaderboard behind models from Google and OpenAI, produced high-quality renderings of natural landscapes from detailed prompts. However, it failed to generate images when given complex instructions involving multiple subjects and scenes, indicating a limitation in handling intricate user requests. Advertising giant WPP is noted as one of the first major enterprise users of the model.
This article is for informational purposes only and does not constitute investment advice.
