sdsds222/Unitale

Implements an LLM-driven pipeline that automatically decomposes scripts into dialogue segments, infers emotional context, and generates character voice descriptions for Qwen3TTS voice design synthesis. The architecture chains multiple AI models—LLM for narrative analysis and emotional inference, IndexTTS/Qwen3TTS for multi-character speech synthesis, and custom heuristics for real-time SFX/BGM insertion timing and scene-aware audio filter application (phone calls, internal monologue, etc.). Runs entirely client-side in the browser with no backend setup required, supporting project serialization via JSON for preservation of audio libraries, filter chains, and script state.