Builds low-latency speech models purpose-built for real-time voice agents, aimed at making an AI voice on a live call sound natural and respond fast enough to hold a real conversation.