Cortexist introduces hybrid Gemma 4 approach combining LLM generation with lightweight JEV-style decision heads for voice turn-taking

Cortexist has introduced a hybrid voice stack that combines Gemma 4 for language generation with a lightweight, JEV-inspired decision head for turn-taking. Instead of relying on a separate model or tag tokens to determine if a reply is necessary, the system uses three tiny MLPs attached to the frozen Gemma 4 backbone (E4B and 12B variants). These heads read the final prompt token’s normalized hidden state from the existing prefill pass and score three actions: speak, defer, or wait.

Performance Benchmarks

On the v0.5 turn-taking set, the local heads outperformed both a zero-shot JEV 1.13.0 call and a Laya pilot trained on the same labels:

• Gemma 4 E4B + head: 95.8% start-turn accuracy, 85.8% interruption accuracy.
• Gemma 4 12B + head: 99.0% start-turn accuracy, 72.6% interruption accuracy.
• JEV 1.13.0: 65.6% start-turn accuracy, 69.8% interruption accuracy.
• Laya: Accuracy in the 30s range.

Against Gemma 4’s own text path on a smaller dev set, the E4B head achieved 90.7% start-turn accuracy compared to 67.4% for the text-only path, and 76.5% interruption accuracy versus 52.9%, all at a fraction of the computational cost of generating tags.

Limitations and Scope

The decision head runs on CPU in under a millisecond once the hidden state is available, resulting in near-zero overhead. However, the current implementation is specialized for multi-speaker voice scenes. Live conversational accuracy drops to 78–85%, and the training labels are synthetic. The system currently over-promotes non-authority turns, with JEV maintaining a lead in authority-priority metrics (93.4% vs. 70.8–81.1%).