Microsoft Announces MAI Realtime 1 Voice-to-Voice AI Model
Microsoft announced MAI Realtime 1, its first voice-to-voice AI model designed for real-time enterprise scenarios with advanced interruption handling. Priced starting at $0.01 per audio minute, the model outperforms competitors like GPT Realtime 2.1 on select benchmarks. This model provides enterprises with a cost-effective, high-performance solution for real-time voice applications such as customer service and interactive training. Its compatibility with the OpenAI Realtime API also makes migration easier for existing developers. MAI Realtime 1 costs $0.01 per audio minute for base usage, with token-based pricing for reasoning tasks, and achieves 98.9% accuracy on the Big Bench Audio benchmark. It has already been adopted by companies like Agoda and Reservations.ai for early trials.
## BACKGROUND
Voice-to-voice models enable bidirectional, full-duplex conversational interactions where AI can process and respond to spoken input natively rather than translating between speech and text. Benchmarks like Big Bench Audio and Tau-Voice evaluate these models on audio reasoning and multi-turn agent task completion.