Improvisation Scriptwriter (Large Language Model Topic Generator): Pioneering Artificial Intelligence–Driven Real-Time Content Suggestion for Engaging Live Video Streams

Abstract

Pavel Malinovskiy*

The dynamic nature of live video streaming demands continuous innovation to maintain viewer engagement, particularly in platforms like Pyjam, where adolescent users broadcast impromptu content such as talent showcases or casual conversations. Traditional streaming lacks mechanisms for real-time guidance, often leading to stagnant flows or lost audience interest.

This paper introduces an improvisation scriptwriter framework, leveraging large language models (LLMs) to continuously generate fresh topics, jokes, or transitions based on ongoing conversation and chat context, providing streamers with engaging prompts to sustain interactive momentum.

Integrated into Pyjam's WebRTC backend, the system analyzes real-time chat messages and audio transcripts (via Vosk STT), employing a local LLM (Llama.cpp-go) for suggestions tailored to teen themes, with latencies reduced to under 500 ms through event-driven triggers and embedding-based relevance checks.

HTML PDF