Sakana AI KAME: Tandem Speech-to-Speech Architecture That Injects LLM Knowledge in Real Time — Accepted at ICASSP 2026

Sakana AI Research·medium signal

Sakana AI introduced KAME (Knowledge-Access Model Extension), a hybrid architecture where a front-end speech-to-speech module (based on Moshi, ~80ms cycle) starts responding immediately while a back-end LLM generates 'oracle' streams of improving responses in parallel. The paradigm shifts from 'think then speak' to 'speak while thinking.' Training data scarcity is addressed via Simulated Oracle Augmentation using synthetic oracle sequences. For builders: this architecture pattern — fast frontend with async knowledge injection from a heavier backend — is applicable beyond speech to any latency-sensitive agent interaction.