Building Offline Voice Agents and Hybrid Edge-Cloud AI Systems

Most voice AI applications today rely heavily on cloud services, but increasingly powerful on-device AI models are opening up new possibilities for developers to build faster, cheaper, more private, and even fully offline voice experiences. In this session, we'll introduce Switchboard, a platform that simplifies building real-time voice and audio applications by making it easy to combine technologies like speech recognition, language models, text-to-speech, duplex voice models, and other audio processing components; some of which run on-device. Through a live demonstration of the Switchboard Editor and a few open-source example projects, we'll explore: Why run AI on-device? The advantages of local processing for cost, latency, privacy, reliability, and offline operation, along with situations where cloud AI still makes sense. Building with Switchboard: How its visual, node-based Editor makes it easy to connect different AI models and audio components, experiment with architectures, and prototype voice applications. Real-world examples: A look at several voice AI implementations, including our Hybrid Voice Agent built for the Future Caribbean buildathon, which combines on-device speech processing and AI with cloud-based intelligence when needed. Getting started: How developers can use Switchboard's open-source examples and templates as starting points for building their own applications, rather than having to develop complex real-time audio infrastructure from scratch. The session will focus on practical demonstrations and real-world applications, with time for questions and discussion throughout.