Change this:
Whisper → Deepgram (nova-2 model) using websockets
TTS-1 → Elevenlabs (Eleven Turbo v2 model) using websockets with optimize_streaming_latency = 3
Use OGG for the audio format.
For GPT-4o, initially get a filler phrase repeating the last thing you asked using GPT-3.5 for faster response and lower latency. Then, obtain the full answer from GPT-4o.