Seems like it could work a lot better still, very quickly, just by merging the trained model with an LLM trained on the language they expect the person to be th

Seems like it could work a lot better still, very quickly, just by merging the trained model with an LLM trained on the language they expect the person to be thinking in. I.e. try to get an equilibrium between the "bottom-up processing" of what the TTS model believes the person "is thinking", and the "top-down processing" of what the grammar model believes the average person "would say next" given all the conversation so far. (Just like a real neocortex!) Come to think, you could even train the LLM with a corpus of the person's own transcribed conversations, if you've got it. Then it'd be serving almost exactly the function of predicting "what that person in particular would say at this point." Maybe you could even find some additional EEG-pad locations that could let you read out the electrical consequences of AMPAR vs NMDAR agonism within the brain; determine from that how much the person is currently relying on their own internal top-down speech model vs using their own internal bottom-up processing to form a weird novel statement they've never thought before; and use this info to weight the level of influence the TTS model has vs the LLM on the output.