The NPCTalk mod, developed by Universal Modder, enables real-time, script-free conversations with non-player characters in GTA 5. Players can interact with NPCs using their microphones, and the characters respond using AI-cloned voices that match their personality and the current game context.
The system integrates several AI models and services: NVIDIA Nemotron ASR for speech recognition, the Qwen 3.8 27B model for text generation, and Fish Audio Drama 3 for voice synthesis. These models run on cloud-based NVIDIA H100 GPUs, while TypeSafe Jev manages NPC animations and physical reactions.
Technical documentation indicates that response latency ranges from 1.15 to 2.2 seconds when using push-to-talk, with an additional 0.7 seconds if continuous monitoring is enabled. The project is open-source under an MIT license, though it requires paid cloud infrastructure to operate.
While currently demonstrated only in the single-player mode of GTA 5, the developer aims to expand this technology to other moddable games. The project was assisted by Claude Code, and the developer has created a reusable agent skill set compatible with various coding assistants to facilitate integration.
