Updates to demo to include audio part

[daavoo]

What's changing

• Add headers and sections to the app.
• Use personal fork and add setup.sh to work around parler_tts installing issues.
• Load models using streamlit cache resources.
  • This prevents reloading on each button click
• Generate and display audio for each chunk.

How to test it

Inside codespaces, run:

bash .github/setup.sh
python -m streamlit run demo/app.py

Additional notes for reviewers

[Kostis-S-Z]

Love this PR! So nice to see the demo working end-to-end, and I like how you split the audio into speaker chunks so you dont have to wait a really long time to get a huge audio chunk file, smart! And great to have the API docs referenced in the UI.

FYI the demo worked pretty fast and smoothly on my 3-year old laptop with no GPU and 16GB ram, max RAM usage was 8GB! I was a bit worried about that since both the text_to_text and the text_to_speech model are loaded simultaneously, instead of sequentially (so that we can offload text_to_text).

Indeed there is this issue with parler not being able to pronounce well certain words, I had missed that in my local testing. But I think we can keep this as it is and merge #26 into #19 and then #19 into main, and change the TTS model from next week.

[daavoo]

But I think we can keep this as it is and merge #26 into #19

Sounds good