But I've read somewhere that KV cache for speech-to-speech model explodes in size with each turn which could make on-device full-duplex S2S unusable except for quick chats.