soft plover
The issue is described by Veeedal here: https://www.youtube.com/watch?v=-s1WCeIthQg
A possible fix is to add attention span + silence detection + hot words.
- The hot word "hey neuro" will activate the attention span value, as long as this value is set high and the AI will assume you are talking to it (input).
- The attention span value will decay to 0 value over time (maybe by cubic).
- Speaking hot words will regenerate the attention span value.
- optional: use vector search to check if the current speech is relevant to previous X inputs. depending on the relevance score add a bit of value to the attention span value
- The hot word "hey chat " disables attention span immediately.
- silence detection will decay the attention span value faster.
optional:
- when attention span is disabled, set the LLM token length generation to a small value at first so it won't try generating long monologues and easier to interject (just like how humans try to feel who wants to lead a discussion)
- gradually increase token length generation as more silence is detected or maybe pick a random chat.
I hope this helps Veeedal :3 peace bro congrats on the debut!
--WIP for now, i still have some commits merge :3
