Smaller Voice Models

In this talk, we’ll cover why some tasks only need smaller models: benefiting from zero cost, privacy, low latency, and offline reliability. Using voice AI as the example, we’ll explore how compact text-to-speech models can run on-device or on-premise, avoiding the cost and latency of cloud-only systems. We’ll share practical lessons from building smaller voice models, where they work well, where larger models are still useful, and how developers can think about model size as a product decision rather than just a benchmark. The session is aimed at anyone building real-time AI applications, voice agents, embedded tools, or privacy-sensitive user experiences.