Decoupled Data Approach for Full-Duplex Dialogue FSMs

To improve turn-taking naturalness in Neural Finite State Machines, this method decouples turn-taking signals from semantic content. By training on real human-human spoken dialogues for timing and human-agent text for semantics, models can achieve full-duplex interaction without the limitations of purely synthetic data.