A phone call, answered on your own CPU
Not one speech-to-speech model but three models taking turns: Moondream's Parakeet Redux hears you, Liquid AI's LFM3-1.7B writes the answer, Gradium's Phonon speaks it while it is still being written. All three are compiled to WebAssembly with fusor-ai and run in this tab; once loaded, nothing you say leaves the page. Every turn is timed from the end of your sentence to the first sound of the reply.