Iconic reveals “Pressure Point”: A showcase for real-time AI digital actors

Interactive studio Iconic launched Pressure Point, a demo showcasing AI “Digital Actors.” Powered by local models, the system enables real-time, natural voice interactions with game characters while keeping gameplay strictly within human-authored narrative bounds

Directed open-endedness and the three-model cascade

The underlying architecture operates on a concept Iconic calls “directed open-endedness,” driven by two human-authored inputs: a persona blueprint defining who the character is, and a narrative blueprint outlining required dramatic beats without hardcoded script lines.

Under the hood, the system executes through a real-time cascade of three distinct AI models:

• Automatic Speech Recognition (ASR):

  • Transcribes the player’s spoken input and analyses emotional cues.

• LLM orchestration:

  • A custom Large Language Model (LLM) engine (utilising Gemma 4) processes narrative context, handles player diversions, and generates the character’s next dramatic beat with embedded emotional markup tokens.

• Contextual Text-to-Speech (CTTS):

  • A proprietary on-device speech engine parses emotional tokens in real time to render natural human delivery—including breaths, pauses, self-corrections, and persistent psychological states.

On-device performance and low-latency execution

A central breakthrough of Pressure Point is its local execution capability, bypassing external cloud data centres to eliminate network lag and minimise operational costs:

• Speech-to-speech latency:

  • The AI digital actor system achieves a median speech-to-speech turnaround time of 1,200 milliseconds (100ms ASR, 700ms LLM processing, and 300ms CTTS generation), with an internal target of bringing latency below one second.

• Hardware efficiency:

  • The entire inference stack runs alongside Unreal Engine on a single consumer-grade GPU.

• Asynchronous scheduling:

  • The pipeline moves non-critical LLM background tasks to moments when the player is silent, allowing instant responses, dynamic player interruptions, and real-time narrative adaptation.

Industry and academic reception of Pressure Point and AI digital actors

Industry researchers and game developers highlighted the demo’s naturalistic dialogue and narrative control.

Professor Julian Togelius (NYU / Unity Advisory Council)

Praised the deployment of local models for removing datacenter round-trip lag and called it the first time he truly enjoyed interacting with an LLM-driven NPC.

Alexandre Moufarek (Director, Google DeepMind)

Noted that pairing conversational fluidity with strict narrative guardrails represents a foundational pillar for future interactive entertainment.

Piotr Mirowski (Senior Staff Research Scientist, Google DeepMind)

Commended the natural push-to-talk voice interface, noting that the mechanical “glitches” feel tied to the character’s internal thinking rather than broken voice synthesis.