Microsoft has unveiled three voice AI models, including one real-time transcription model and two speech synthesis models. The streaming transcription model converts speech to text in real time and updates transcripts as users speak. It delivers an initial result in an average of 320 milliseconds, supports more than 60 languages and costs $0.54 per audio hour. The company also highlighted a cheaper non-streaming transcription model released last month.

With memory supply shortages raising costs for DRAM-heavy wearables, AI device makers are cutting per-device DRAM through code optimisation, Qualcomm said. Qualcomm is also supporting 1-bit models on its Snapdragon AR1 line to reduce memory use further. Executive vice president Ziad Asghar said smaller devices face tighter memory limits and could see lower component costs. Qualcomm demonstrated a 1-bit multimodal model for AI glasses and described moves toward lighter platforms in earbuds.