GTA V NPCs Respond in Under 2 Seconds With New AI Mod NPCTalk

Universal Modder has launched NPCTalk, an open source mod for Grand Theft Auto V, which enables players to have real-time conversations with non-player characters (NPCs). NPCTalk transforms the game into a real-time conversational experience where scripted dialog is replaced by a more contextual and dynamic dialog system. Players can control NPC behavior with verbal communication. The launch of NPCTalk is significant because it highlights the consumer accessibility of AI voice technology for gaming applications.

Universal Modder launches open source tool for real-time NPC conversations

NPCTalk leverages three different AI models to orchestrate the conversation process. First, NVIDIA Nemotron ASR converts players' voice input into text. Second, NPC responses are created by Qwen 3.8 27B deployed on Cerebras' infrastructure. Third, Fish Audio Drama 3 is used to produce the NPC's text response as a voice clone of the specific character.

The system uses a cloud-based inference model, utilizing an NVIDIA H100 GPU, to handle the complex computational demands. Response times differ depending on the mode of input. For push-to-talk, the response time can be between 1.15 and 2.2 seconds. When enabling open microphone detection, an additional 0.7 seconds can be added to the conversation loop.

According to Universal Modder, NPC responses take less than a second. Although the published push-to-talk response time indicates up to 2.2 seconds. Universal Modder announced the effort in a public statement.

The mission was to make all the NPCs respond like the character using the cloned voice for each NPC. The toolkit is accessible on GitHub with an MIT license, so developers can see the code.

A free and playable version of the GTA V mod will be released next week. The GTA V mod is free, but hosting the cloud services and GPU infrastructure needed may cost money if the user has a custom instance. This is an important distinction between the consumer version of the mod and the technical backend.