Some feedback - the primary problem with your current tool chain that makes using the product a - wow "ai" is annoying experience instead of - wow this is cool - is the inability of the LLM/TTS response to dynamically assess the appropriate amount of time to wait to respond. You've got these parameters which is a good start. But enabling the user on the call to dynamically adjust these parameters with a tool call that is triggered by phrases like - please stop interrupting me. Would be a good start.
In the long run, a speech to speech model trained on natural human conversation is the best bet here. But in the mean time, giving users control in natural language is a start. When I think about whether or not I want to implement the tool as a sort of sales qualification screening for training LLMs models for customers, the number 1 reason I don't is most prospects don't seem to know what to do/say when talking to the bot. And they seem to have a bad enough experience that it's more likely to turn the customer off from using AI than it is to make the customer interested in signing up based on the customized/1:1 responses they get.
https://cdn.discordapp.com/attachments/1211486158232617032/1274395247383806003/image.png?ex=66c21890&is=66c0c710&hm=35b83cb5d7c888406ac8068ce03a8e26e00db89389647420225baacfc5107eb6&