I disagree on the voice interaction point. I DO want voice modes to be glorified STT-TTS.

The OpenAI AVM is uncanny and over-eager to me. You can run Anthropic voice on Opus since a handful days and it's incredibly snappy without feeling weird.