We built this example as part of a buildathon. It is a real-time voice agent that runs fully offline on iOS (but React Native based so Android will come soon).
VAD, ASR, a small language model, and TTS all run locally on iOS, with a cross-platform architecture that can swap the local model for a frontier cloud LLM when needed.
For now, it's just a button that lets you switch between the on-device SLM and the cloud-based LLM. But we're planning to implement different routing options (e.g. automatic fallback when internet is available, or confidence based, etc.).
Would love to see this remixed. It's designed to be easy to swap components (e.g. use a different VAD, STT, TTS, different language models, etc.).
The implementation is open source. You'll see there's a dependency on the Switchboard SDK. Currently, the free tier is only to 10K MAUs, but we're planning to remove that limit and make most of the core offering a totally free service -- so get in touch if the current model is a blocker for some reason.
jimsrand•22m ago
VAD, ASR, a small language model, and TTS all run locally on iOS, with a cross-platform architecture that can swap the local model for a frontier cloud LLM when needed.
For now, it's just a button that lets you switch between the on-device SLM and the cloud-based LLM. But we're planning to implement different routing options (e.g. automatic fallback when internet is available, or confidence based, etc.).
Would love to see this remixed. It's designed to be easy to swap components (e.g. use a different VAD, STT, TTS, different language models, etc.).
The implementation is open source. You'll see there's a dependency on the Switchboard SDK. Currently, the free tier is only to 10K MAUs, but we're planning to remove that limit and make most of the core offering a totally free service -- so get in touch if the current model is a blocker for some reason.