1. Kokoro-82M (In-Browser TTS)
• Where it runs: 100% inside the browser on device via WebGPU / WebAssembly.
• Model size: ~85 MB ONNX weights.
• Purpose: Real-time voice for dynamic player-typed inputs ("Carve Your Own Path") with $0 API cost and no backend dependency.
2. StyleTTS2 (Diffusion Model)
• Where it runs: Offline in Python/PyTorch (local GPU or Google Colab).
• Purpose: Extracts a style vector from a 3-second reference audio to mimic wild cartoon prosody.
• Connection: Any clip generated here can be passed as reference audio to scripts/test_styletts2_voice.py!
3. Curated 40 Soundboard
• Where it runs: Pre-rendered audio sprite sheet (Web Audio API).
• Purpose: 0 ms instant latency for fixed game dialogue ("NARF!", "POIT!", "Take over the world!").
• Audition below: Test all 40 clips with full procedural synthesis fallback!