POC for voice question/answers

[lfoppiano]

Let's write a Streamlit app that can be used to test the end to end Q/A with voice (no document viewer) but a transcript of what is being requested / responded

• Deploy everything (STT, TTS) on modal.
• Real time voice generation
• Play back what was transcrabited
• Merge both models(TTs and STT) in one container
• Remove the buttons VAD (Voice Activity Detection)
• Integrate LLM for chat
• Deploy to HF space
• Integrate the llm engine with the qa library

[lfoppiano]

Upon testing, there is an issue to spin up the modal.com synthetizer.

A few comments:

• we should make a sort of request once someone upload a PDF file, so that the model start warming up.
• ⏹Click the mic to start is misleading, just remove it -- I spent 10 minutes clicking on it before figuring out it was down -- I'm getting old perhaps...
• the document panel is too big, we should not need to scroll down, remove chunked size and context window, the default will be sufficient
• when I try to record I always get "connection failed", unfortunately I cannot access the model
• I notice the huggingface space have only secrets as env variables, unless they are secrets, like password or api token, should be visible, so that we can update them without overwrite them blindly